FIX HGBT exclude zero-weight samples in tree grower - #27

Reference Issues/PRs

Addresses the zero-weight / deep-tree failure discussed on scikit-learn#34838 and scikit-learn#34745.

What does this implement/fix?

Root cause: HGBT split constraints use unweighted histogram counts (hist.count). Zero-weight rows still entered the grower partition, so they inflated min_samples_leaf / split-stopping even though their gradients and hessians were already 0. With min_samples_leaf=1, deeper trees could keep splitting nodes that would already be leaves if weight-0 rows were absent.

Fix: Pass sample_weight into TreeGrower / Splitter and initialize partition with positively weighted sample indices only (same approach as DecisionTree*). Zero-weight samples never enter a leaf, so they cannot form a zero-weight leaf and no longer affect unweighted count constraints.

Removed the previous fit-time pre-filter / integer-weight expansion helper.

Verification:

• test_sample_weight_zero_equivalence_deep_trees and grower unit test for partition exclusion
• Dense common check enabled with min_samples_leaf=1, max_depth=1 (deeper trees can still diverge for positive integer weights due to float32 weight * grad vs summing unweighted gradients)

Still limited: min_samples_leaf > 1 with non-unit positive weights can still diverge from sample repetition until leaf-size constraints use weight sums.

AI usage disclosure

This pull request includes code written with the assistance of AI.
The code has not yet been reviewed by a human.