WIP: still being benchmarked/calibrated/refined

Reference Issues/PRs

Implements items 4, 5 and 6 of #34764 (comment)

Closes #30662

And I'd say it also closes #14306. Though this issues does have some interesting idea that are not implemented here, especially parallelism over block of samples x features for the histogram building. Still, benchmarks will (I hope) show HGB does continue to scale until 16-32 threads for many medium/big datasets with this PR. I personally would consider that good enough ^^

What does this implement/fix? Explain your changes.

• "use_threads_if" heuristic: TODO
  • Note: I initially tried dynamic sizing of the thread pool but it is catastrophic: OMP apparently recreates a thread pool each time it's resized.
• Limitation of the initial thread pool size: in some cases (small datasets, small number of features), we know from the start a big number of threads will not help any parallel region. So we down size it to make sure
• Cached calibration of how much overhead a prange represents for a given number of threads.

AI usage disclosure

I used AI assistance for:

• Code generation: closely guided
• Test/benchmark generation
• Documentation (comments)
• Research and understanding: a lot

Any other comments?

Benchmarks on the way.