TTreeView's internal logic will be changed by future commits to play
nice with TBB tasks, and will not be able to provide the interface that
this logic was relying on.

Each TDF node keeps a thread-local tuple of TColumnValues. TColumnValues
read the values of a column for a certain entry, but entry ranges and
TTreeReader-related objects are _task_ local.

Before this change, in case of interleaved task execution, the second
task starting in a thread would re-initialize the TColumnValues and
 overwrite the previous task's settings.

Now TColumnValues deal with multiple initializations by keeping their
internal state in a stack: when a new task initializes a TColumnValue
it pushes the required values in the stack, when a task completes
execution it pops its values from the stack.

Note that interleaved TBB tasks always follow a last-to-start,
first-to-finish pattern, which allows us to use a LIFO structure for
TColumnValue.

This fixes a race condition in which a TTreeReader and its
TTreeReaderValues could be deleted concurrently:
Thread #1) a task ends and pushes back processing slot
Thread #2) a task starts and overwrites thread-local TTreeReaderValues
Thread #1) first task deletes TTreeReader

Increase the fit tolerance for the unbinned log-likelihood fit test. This should fix some failures observed in MT mode

Do not use std::accumulate in the reduction function for the LogLikelihood to be sure the correct order is used

Update zlib to 1.2.8 released in Apr 2013. This is a battle tested
version which is used as a base for QAT and Cloudflare zlib forks.

Signed-off-by: David Abdurachmanov <David.Abdurachmanov@cern.ch>

Vc only supports x86_64, but the previous condition allowed it to be
enabled on x86 systems, breaking the build.

This property isn't available in the clang-format we use, so
travis on github fails parsing this configuration. It also seems
to suggest now a default code style instead, so there's a chance
the last few commits have the wrong formatting now.

Fix a big in the IntegralEvaluator class used for computing the bin integral when fitting in case of Veccore types and multi-dimensional functions. Fix also the padding in BinData when the number of points is a multiple of the vector size.

Due to a previous change in the gradient boosted DT the parameters
used in the stressTMVA test case was just below threshold.

This change modifies the shrinkage to be 0.1 instead of 0.3. Both
choices are reasoable "black-box" parameters meaning it's a very
limited tuning.

the optimisation level is too low. The test takes more than 20 minutea to
run therewith triggering a timeout on arm nodes systematically.

With a little help from Chris Jones - thanks!

roots tests are fine.