The `RooRealIntegral` class is smart enough to figure out which
variables the function the integrated function doesn't depend on and
trivially integrates them itself by multiplying with the variable
definition range.

However, if the integration range is a subrange of the variable range,
this was not considered. This resulted in wrong results. for integrals
like `pdf.createIntegral(x, "subrange")`, where the pdf doesn't depend
on x. These kind of integrals can occur in the projections that the
RooAddPdf does, so it's important that they work, and fixing this
partially addresses #11486.

This change also fixes a so-far unknown bug in the `RooXYChi2Var`, which
also used these kind of integrals. Without this fix, the `Integrate()`
feature for `chi2FitTo()` was completely broken, which can be seen in
the output of the `rf609` tutorial with any previous ROOT version. The
tutorial looks okay by chance, because the function is dominted by the
quadratic term in `x` that is constant in the fit. But if one makes this
a floating parameter, the problem gets obvious.

Probably that was the reason why the main model parameter was set
constant to begin with, to sweep the bug under the rug. Now, the
tutorials are updated to have the quadratic coefficient floating too.
And also `stressRooFit`, since the reference file has to be updated
anyway.

To demonstrate that things work correctly now, a new unit test was
implemented that does the closure check of the `integrate()` feature of
the `RooXYChi2Var` with a linear fit function.

In the `rf205` tutorial and the associated stressRooFit test, the `bkg`
RooAddPdf was created using the `sig1frac`, while it is clearly the idea
to use the `bkg1frac` parameter.

This is fixed in this commit, also fixing the name of the `bkg1frac`
parameter, which was accidentally set to `sig1frac` (probably a
copy-paste error).

However, to not change the output of the tutorial and the stressRooFit
test, the value of `bkg1frac` was set to the same value as `sig1frac`,
such that this commit only fixes the model but doesn't change the
tutorial and test output.

The `rf901_numintconfig` tutorials and related stressRooFit test use the
RooLandau class to demonstrate numeric integration, because it didn't
support analytic integration. Now that it does, the analytic integration
needs to be disabled with `forceNumInt`.

In the WrappedMultiTF1 when trying to compute the Hessian for linear functions via TFormula::HessianPar , the code asserts in CLAD since there is no Hessian to compute.
Return then by default a zero value

[math][minuit2] Fix some warning message in Minuit2

[math][fit] Enable External Hessian computation only for Minuit2 to avoid Hessian check in case of linear fitting

Fix also an issue with TFumiliMinimizer (divide correctly by  2 in chi2)
and fix a bug in computing Chi2 hessian element that is used by fumili

In stressHistoFit uses Fumili2 for test without analytical gradients (w/o option G)

In the implementation of `RooHistPdf::plotSamplingHint`, some of the
returned values can be outside the variable definition range. This might
lead to invalid evaluations when the sampling hints are also used to
evaluate other PDF components of a composite PDF, as reported in #11482.

This commit suggests an improved algorithm for the sampling hint:
consider only bin boundaries that are strictly between the plot limits
within some epsilon, and for the limits just sample two additional
points left and light that are some epsilon aways from the plot limits.

The `stressRooFit_ref.root` reference file needs to be updated with this
change, because the sampling changes the RooPlots that are used as
reference objects.

Closes #11482.

The `stressRooStats` tests are now run also with the different BatchMode
backends as unit tests.

Furthermore, the backend for the default tests is explicitly set to
BatchMode "off", such that we still keep testing the legacy RooFit
should the BatchMde become the default one day.

This is already done in `stressRooFit`, and it's useful because it gives
more flexibility in executing the test.

The problem was that the previous `$ROOTSYS/test/*_ref.root` path
required ROOT to be installed, which is not always wanted for developers
that just like to run the tests.

The test currently fails with
-12823.35859557967 -12823.35859557937 -12823.35859557937
Test DeltaR failed !!

This is to suppress RooFits warnings.

This commit contains automatic changes done by `clang-tidy`.

Create a new type of merging, a special case when histogram has an axis with label and one axis can be merge numerically but are compatible.

A new test is added also in stressHIstogram for this new case

This should fix ROOT-10828

There still was a constructor of `RooDataSet` that was hardcoded to use
the TreeDataStore. This needs to be changed, because the
RooVectorDataStore is more performant and compatible with the new
BatchMode.

In the `foldIntegrals()` function that reverses the unfolting of
normalization integrals in the computation graph, some custom logic was
used to redirect the servers. However, it caused some failures in the
ASAN builds because it accessed invalid memory.

This commit suggests to fix that by using the already existing
`RooAbsArg::recursiveRedirectServers()` function for the same logic.

Furthermore, in `unfoldIntegrals()`, clients outside the computation
graph are not considered for the server redirection, which also fixes
ASAN build failures.

All found by Clang's -Wunused-but-set-variable.

In stressRooFit, you could already globally select the BatchMode for
`fitTo` calls. With this commit, the selection is also considered for
`createNLL` calls, widening the test coverage of the new BatchMode.

In particular, this tests the plotting of likelihood functions, which is
much faster when using the new BatchMode backend (up to a factor of 10
in the first profile NLL tuutorial).

It is available less and less often, and we do not actually
benefit a lot from printing the distro (stress) or we can get
the same info from /etc/os-release (cmake).

This is to suppress the warnings we see now in the nightlies on the
macbeta nodes:

https://lcgapp-services.cern.ch/root-jenkins/view/ROOT%20Nightly/job/root-nightly-master/LABEL=macbeta,SPEC=cxx17,V=master/lastBuild/parsed_console/

After this PR, there will only be a final PR necessacy to fix these
warnings in TMVA.

Before, the interpreted tests were not declared as longtests, and the
compiled versions were not. This is now flipped, such that the compiled
BatchMode CPU and CUDA tests are also run without `longtest`.

With some recent updates to the RooFit BatchMode, it can now also be
activated for the RooMCStudies in `stressRooFit_tests` as the fit
results are the same.

The direct creation of RooFit test statistics instances is discouraged
because the more configurable `createNLL` and `createChi2` should be
prefered.

In the RooChi2Var, there was another case of inconsistency between
constructors that seemingly do the same but actually don't.

A RooChi2Var should be created with the constructors that take RooFit
command arguments. However, there were also other constructors used by
the RooFit test statistic implementation details (i.e. in
`RooAbsOptTestStatistic::create`) that when used with default aruments
behave inconsistently with the command arg constructors: the default
error type is different, and errors will be estimated from the pdf and
not from the data.

This lead to the confusing situation that when creates a `RooChi2Var`
without any command arguments, the default error mode is suddenly
different. This inconsistency should be removed by having only the
command argument constructors part of the public interface.

Making these constructors private should not be a problem. There were
already considered implementaiton details before, and their interface
was already changed anyway to use the configuration structs in 6.26.
Nobody has complained about that so far.

Closes #10557.

Part of the RooFit code modernization, avoiding the use of unnecessary
wrapper classes and using a simple templated struct instead.

The test fails (access violation) on Windows 64 built in RelWithDebInfo mode when running with `ctest` (i.e. it runs fine from the command prompt). So let's disable it on win64 for the time being.