PRF: Optimize binning in HGBT - #34248

This PR optimizes binning in histogram gradient boosting, specifically _find_binning_thresholds.

Closes #34304

Initial motivation

The weighted path is currently very slow for datasets with fewer than 200k samples because _weighted_percentile dominates the binning time. For larger datasets, sample weights are consumed by subsampling, so this bottleneck does not apply in the same way.

For example, on a 100k x 32 weighted dataset, binning takes more around 90% of the total fit time. The main issue is that the current implementation loops over percentiles and calls _weighted_percentile once per percentile. A single call to _weighted_percentile is still not that efficient, especially with many threads, so this PR adds a small local helper optimized for this use case.

What does this implement/fix? Explain your changes.

This PR optimizes threshold computation by:

• using a local _weighted_percentile_1d_sorted helper instead of _weighted_percentile;
• using np.sort instead of np.argsort for the unweighted case;
• avoiding a mask allocation for NaNs in the unweighted case by using np.searchsorted;
• replacing np.unique with a simple distinct-value mask on already sorted data.

AI usage disclosure

I used AI assistance for:

• test and benchmark generation.

Speed-up

Weighted datasets under 200k samples with more than 255 distinct values get the largest speed-ups. For small datasets the speed-up can be extremely large, and for a typical 100k x 32 dataset it is already around 100x.

For unweighted datasets:

• single-threaded execution is typically 2-3x faster;
• multi-threaded execution can be up to 2x faster, although the gains are often smaller.

Detailed benchmark results were produced with this benchmark script and this runner script. The tables report speed-up of this branch relative to main.

On a large machine:

On my laptop: