Redefinition of a custom column is currently allowed due to lack
of checking and sort of works (last definition "wins"), but will
be checked for and will count as an error in the future.

Actually, let's remove GetCustomColumns altogether and move the logic
in TInterface.

This name should make it clear that the method returns the actual
graph nodes, and not the names of the defined columns.

Since the test script is not a fully compliant C++ file, clang-format can not
(be expected to) follow its unusual conventions/requirements.

Fix registration of other atexit functions during an atexit handler. Recursive registration of atexit handlers is legal and should be handled, not ignored.

Hide members of CXAAtExitElement and make it a callable object. Update iteration to use llvm::reverse and C++11 range based for loop.

See https://sft.its.cern.ch/jira/browse/ROOT-8969

We will be reading the tree, so make sure that the user
did not forget to remove pointers to out-of-scope vars.

The user should *always* call ResetBranchAddresses on their trees
when they are done with them, but in case they forgot this can save
some debugging time at no cost.

A TTaskGroup represents concurrent execution of a group of tasks.
Tasks may be dynamically added to the group as it is executing.

This is a breaking interface change: Filter now returns templated
TInterface<TFilter<F,P>> objects instead of TInterface<TFilterBase>
as before (and anagolously for Define and Range).
Jitted transformations still return TInterfaces wrapping the base
classes (e.g. TFilterBase).

Filter, Define and Range returned TInterfaces wrapping pointers to the
base classes representing the corresponding nodes. This choice was made
to limit compile time and ease the introduction of jitted transformations.
As a consequence, nodes of the functional graph communicated with each
other through virtual calls.

This commit lets nodes call into each other through the derived (template)
types instead. Given that all nodes' virtual methods are marked `final`,
this should allow compilers to de-virtualize such calls.

We can't use the interpreter when generating a PCM as this would
generate AST nodes which then would end up in the module, which
is causing a long chain of modules (such as redefinitons as we
suddenly have the same cling warpper function multiple times).

In this code path we seem to always have a number that we want
to convert to a string. So let's just use strtol instead here if
the argument is just a number, which should avoid the issue with
the generated code.

As we also now check if the input is a number, I added an assert
that verifies we fail when we can't handle the given user input.

Certainly the ownership is quite incomplete - but should provide a
sufficient starting point.