This is done such that when importing the JSON back, names are not
unexpectedly changes.

This is to be consistent with the old HistFactory implementation.

For the model parameters, the HS3 HistFactory implementation was able to
generate them implicitly if they are not in the JSON. Now this is done
in a consistent way also for the observables.

So far, this was done in a differnt way that re-used the internals of
the RooJSONFactoryWSTool.

Like this, it is easier to extend the test because the ROOT files
doens't have to updated all the time.

  * round numbers for Barlow-Beeston reverse-engineering
  * sort normalization factors
  * harmonize names of shape systematics

Improve the binomial fit use by TEfficiency::Fit by setting the correct chi2 and number of degree of freedom.

Add function in Fitter to set number of fit points and fit type. When the fit type is a binned likelihood fit used as chi2 value to report 2 * minFCN

Fix also the copying of the TF1 object after fitting. Use IsA()->New() instead of copy constructor which will not work correctly for TF2 and TF3 objects

This follows up on a previous commit 92fbd0db with the title:
```
[RF] Modernize `RooVectorDataStore::RealFullVector` class
```

It changed the type of the `RooVectorDataStore::RealFullVector` members from
`std::vector<double>*` to `std::unique_ptr<std::vector<double>>` while
increasing the class version.

However, while reasing old workspaces with the previous class version
worked perfectly fine, there were logged errors when writing and then
reading back a dataset with weight errors in ROOT `master`:

```
TStreamerInfo::BuildOld:0: RuntimeWarning: Cannot convert RooVectorDataStore::RealFullVector::_vecE from type: vector<double> to type: vector<double>, skip element
TStreamerInfo::BuildOld:0: RuntimeWarning: Cannot convert RooVectorDataStore::RealFullVector::_vecEL from type: vector<double> to type: vector<double>, skip element
TStreamerInfo::BuildOld:0: RuntimeWarning: Cannot convert RooVectorDataStore::RealFullVector::_vecEH from type: vector<double> to type: vector<double>, skip element
```

This was very unexpected. Maybe there is some problem with the IO of
class members of type `std::unique_ptr<std::vector<double>>`. This
commit suggests to fix this problem by using `std::vector<double>`
directly, which makes the code much simpler anyway. It also allows us to
get rid of a workaround in `RealFullVector::Streamer()` where it was not
clear if this workaround is still needed or not.

The class version is not incremented further, because the last increment
only happened a few days ago, and no new ROOT version has been released
since then.

The RooHistPdf and RooHistFunc should be able to do analytic integration
if the input is a linear transformation of a variable using
RooLinearVar.

This makes the fits faster where one wants to shift a template on the
x-axis, which is for example talked about in this forum post:
https://root-forum.cern.ch/t/roofit-pdf-normalization-integration/53905

The last normalization set should not be stored by pointer because this
can result is false positives when comparing normalization sets by
pointer: a new normalization set might have the same address as the
previous one.

The evaluateSpan() function was superseeded by computeBatch().

* The option TEXT90 did not work together with COL

* 90 is possible

Exporting the dependants invalidated the `elem` iterator, so it can't be
used anymore after calling `exportDependants()`.

It is sufficient to write the axes once for each HistFactory pdf instead
of repeating them for each sample.

Use the replacement TYPED_TEST_SUITE, and have a compatibility define
for gtest version < 1.10.0.

* Fix an issue with the canvas editor when the canvas is embedded

Fix a problem with the editor not appearing when the TCanvas is embedded, [as reported on the Forum](https://root-forum.cern.ch/t/editor-not-appearing-in-gui-canvas/53968

)

Co-authored-by: Axel Naumann <Axel.Naumann@cern.ch>

* Cosmetics (thanks Axel!)

---------

Co-authored-by: Axel Naumann <Axel.Naumann@cern.ch>

Add documentation for Vary regarding runtime that does not scale linearly with the number of variations

This singleton helper class should not be serializable.

It was not good to have this signature in RooAbsData, because the
implementations in RooDataHist and RooDataSet were inconsistent.

The RooDataSet indeed took the weight error as the third argument, but
the RooDataHist version instead took the sum of weights squared, which
is equivalent to the squared weight error.

That means using `RooAbsData::add(row, weight, weightError)` results in
different weight uncertainties depending on which data implementation
you use. This should not happen. It probably didn't happen to many
people, but I was affected by this when implementing
`RooAbsData::split()`.

The solution to this problem it to remove the `add(row, weight,
weightError)` from the virtual RooAbsData interface. The `add(row,
weight)` signature can stay, because this one is implemented
consistently in RooDataSet and RooDataHist.

The file was already partially formatted correctly, and it is now
formatted completely because the next commit is about to touch the file.