Small optimization for block hist matrix.

[trivialfis]

This is an early WIP. Please ignore copied and pasted code. Initial benchmark with HIGGS:

• Platform : Ubuntu 18.10
• CPU : AMD Ryzen 7 2700 Eight-Core Processor

Results:

• Time spent in GHistIndexBlockMatrix::Init

Before	Now
7.56465 s	4.09427 s (-45.87%)

• High water marked (peak) memory usage with GHistIndexBlockMatrix::Init:

	Before	Now
Child high-water RSS	12138724 KiB	8359052 KiB (-31.13%)
Recursive and acc. high-water RSS+CACHE	19922500 KiB	8371424 KiB (-57.98%)

This is mostly for addressing memory usage issue. The benchmark is carried out by tests/benchmark/block_hist_init.cc. Hopefully with more effort we can bring back feature grouping, which might be beneficial to GPU Hist too.

One particular thing worth noting is that, I tried injecting OMP into many places, it ends up single thread version is the fastest one.

[trivialfis]

Sadly this is not even close to fixing #2326 .

[trivialfis]

@hcho3 The benchmarking code is not necessary, but very useful for #4503 . You can decide whether we should keep it here.

[hcho3]

@trivialfis I'll review this PR later in detail. I think we should keep the benchmark code. Later, we can expand it as part of a regular performance benchmark.

[trivialfis]

@hcho3 Take your time. This is part of memory optimization for hist and is not the most critical one.