The rule were actually still executed on the proper object and
in the right order.  However when we had

  - a top level object with a rule for one of its transient member
  - that top level object also having a data member that is an STL collection.

then the fOnfileObject (type TVirtualArray) was inadvertently resized
by the top node of the STL object (i.e 'type 4' branch) to the size of
the collection.

Before the improvement to the rule scheduling the ordering of operations
was:
   1 - a - resize the top level object's fOnfileObject to 1
   2 - b - read the input for the transient member into the fOnfileObject
   3 - c - execute the rule using the input/fOnfileObject and updating the transient member
   4 - d - resize the top level object's fOnfileObject to size of sub-collection

The last stage provoke a reallocation of the element of fOnfileObject (thus
losing the input but only 'after' it has been used.

With the new rule scheduling the order becomes

   1 - a - resize the top level object's fOnfileObject to 1
   2 - b - read the input for the transient member into the fOnfileObject
   3 - d - resize the top level object's fOnfileObject to size of sub-collection
   4 - c - execute the rule using the input/fOnfileObject and updating the transient member

and now the consequence of the inadvertent resize becomes visible.
i.e. for each entry where the sub-collection 'grows' the transient member
value will be invalid. [Note: in the example provided with ROOT-9701 the
sub-collection 'grows' by one at each entry]

Fwd decls from rootmap files were injected to ROOT at the startup time.
However with modules we don't need any foward decls as we're preloading
all modules.

Our future plan is to remove rootmap files as a whole, but we still need
to support more features (see 2380). So for now we can remove what we
are not using.

This improves startup memory by 10 mbytes.

This reverts commit 1a24e092.

Fixes https://sft.its.cern.ch/jira/projects/ROOT/issues/ROOT-9624

Description: https://bugs.mysql.com/bug.php?id=85131

NetxNG needs the include directory for XRootD, but when XRootD is
enabled as a builtin, the include directory points to the build
directory and causes CMake to fail with an error like the one below:

CMake Error in net/netxng/CMakeLists.txt:
  Target "NetxNG" INTERFACE_INCLUDE_DIRECTORIES property contains path:

    "${BUILDDIR}/include/xrootd"

which is prefixed in the build directory.

This is a blocker to move to modern CMake constructs (using targets
and properties) for targets that use headers of builtins in their own
headers, as we cannot create imported or exported targets with proper
dependencies.

of branches when calling the TTree::Get(Tot|Zip)Bytes method.

Commit b6523df2 introduces the infinite recursion whenever
there was a problem reading the StreamerInfo record and
introduces a useless attempt to read from file in the case
of a class derived from TFile that overrid GetStreamerInfoList.

To allow for increase run-time performance and increase thread
scalability the override ability of TFile::GetStreamerInfoList
is replaced by an override of TFile::GetStreamerInfoListImp with
updated return type and arguments.   If a class override
TFile::GetStreamerInfoList you will now see a compilation error like:

```
/opt/build/root_builds/rootcling.cmake/include/TSQLFile.h:225:19: error: declaration of 'GetStreamerInfoList' overrides a 'final' function
   virtual TList *GetStreamerInfoList();
                  ^
/opt/build/root_builds/rootcling.cmake/include/TFile.h:231:24: note: overridden virtual function is here
   virtual TList      *GetStreamerInfoList() final; // Note: to override behavior, please override GetStreamerInfoListImpl
                       ^
```

Instead you need to override the protected method:

```
   InfoListRet GetStreamerInfoListImpl(bool lookupSICache);
```

which can be implemented as

```
  InfoListRet DerivedClass::GetStreamerInfoListImpl(bool /*lookupSICache*/) {
       ROOT::Internal::RConcurrentHashColl::HashValue hash;
       TList *infolist = nullptr;
       //
       // Body of the former Derived::GetStreamerInfoList with the
       // return statement replaced with something like:

       // The second element indicates success or failure of the load.
       // (i.e. {nullptr, 0, hash} indicates the list has already been processed
       //  {nullptr, 1, hash} indicates the list failed to be loaded
       return {infolist, 0, hash};
  }
```

See TFile::GetStreamerInfoListImpl implementation for an example on how to implement the caching.

We assumed that we will always have module file and unconditionally disable
header parsing on demand. However, the major use-case is gradual migration
to modules. In this scenario (tested by root-meta-fwdDecls-fwdDeclarations),
we have a dictionary which has no module file and still relies on the old
behavior. This can realistically happen when users gradually migrate to
modules. For example, we have modules-aware ROOT and untouched third party
code.

This patch enables header parsing on demand when we have no module file
available.

Now these are: RWebWindow and RWebWindowsManager

This fixes ROOT-9693 and ROOT-9703