Issues reported by coverity:
 - https://coverity.cern.ch/query/defects.htm?projectId=10001&cid=98864
 - https://coverity.cern.ch/query/defects.htm?projectId=10001&cid=98876

This is part of 4ff4d9aa.

A plain include_directories() was meant to be converted
into target_include_directories() in the previous commit,
but the second part was missing.

and use a simple vector of pointers.

When noting that a TStreamerInfo set has already been processed, we
must do it in 3 separates steps:
  - check it was seen before
  - read and process the set
  - record that the set as been seen.

The previous situation:
  - check and record that the set as been seen
  - read and process the set
led to a race condition if a second thread was checking the same
set before the second step was completed (in which case the
second thread was believing that the set was process and looking
for the result of the process (one of the StreamerInfo) but could
not find it.

We extend RConcurrentHashColl to have 2 new operations
  - standalone Hash calculation
  - standalone Find of hash
  - standalone Insert of hash.
and we use it to split the check and the recording as described
previously.

(this is an addendum to 95bf468438)

The new release includes some improvements in both Forward and
Reverse mode:
  * Better correctness of C++ constructs -- handle scopes properly; allow proper
  variable shadowing; and preserve namespaces.
  * Efficient evaluation in forward mode.
  * Reduced cloning complexity.
  * Handle more C++ constructs -- variable reassignments and for loops.

See more at: https://github.com/vgvassilev/clad/blob/v0.3/docs/ReleaseNotes.md

[TMVA] Configure tutorials/tmva/makefile

Example makefile from tutorials/tmva should use the same c++ standard as
ROOT (otherwise one faces errors from string views in the root headers
when building the tutorials).

The header TProcessExecutor.hxx is from core/multiproc.
It was being removed before, so there was no problem if
didn't exist, but since the logic has been inverted to
add if needed, it created a problem.

The `PrintMatrix` was declared in the global namespace which means that
if you define your own PrintMatrix macro you can have a name clash.

This PR "qualifies" the name using the prefix TMVA_DNN_

In MethodDNN one could start training with a batch size larger than the
number of events in training leading to non-sensical output. TMVA now
properly warns (fatal warning) and suggests a fix (decrease batch size).

Explicitly use bools in if's
Use `.empty()` instead of `.size() == 0`

Simple to //-ise over number of output classes. Doing the same over
number of trees seems to yield no benefit for ~1000 depth=3 tree so this
was skipped (binary and multi class).

Optimal would be to //-ise over events, but the current design makes
this tricky.

The INCLUDE_DIRECTORIES property only became available in
CMake 3.11, so we have to use COMPILE_FLAGS instead, which
is already available in our required version (CMake 3.4.3).

Note: The ROOT_ARCHITECTURE value of win32old is not found anywhere
else in the code, and the files referred to don't exist anymore, so
the obsolete linkdef file has been removed too.