This project provides several fast $O(n\log n)$ implementations of weighted quantile-based splits (pinball loss) for decision tree regression.

It replaces the previous $O(n^2)$ absolute-error path in scikit-learn (see PR #32100), reducing training time by hundreds of times on large datasets.

• Two-heaps algorithm: expected $O(n\log n)$, fast and simple to grasp.
• Segment tree / Fenwick tree variants: guaranteed $O(n\log n)$ worst case; slower but theoretically clean.
• Supports arbitrary quantiles ($\alpha\in[0,1]$) and weighted samples.
• Implemented in pure Python + Numba, easy to integrate or adapt. See scikit-learn PR for a Cython implementation of the two-heaps algorithm.

Full technical report: report.ipynb