Cut head_tail_breaks and box_plot dask re-scans

[brendancol]

Summary

• _run_dask_head_tail_breaks: persist data_clean once, track the running mask count across iterations, and fuse the mean and head-count reductions into a single dask.compute() call per iteration. Cuts per-iteration graph traversals from 3 to 1 and eliminates the re-read on every loop pass.
• _run_dask_box_plot (new) and _run_dask_cupy_box_plot: replace data_clean[da.isfinite(data_clean)] (which forces compute_chunk_sizes) with the same seeded _generate_sample_indices sampler that natural_breaks and quantile already use. Percentiles are then computed on the finite portion of the sample in numpy.

Motivation

Static analysis flagged three HIGH-severity patterns on the dask backends of classify:

1. _run_dask_head_tail_breaks ran .compute() inside a while loop for the mean, new-mask count, and total-mask count — 3 full graph traversals per iteration, N+1 iterations typical.
2. _run_box_plot(..., module=da) used boolean fancy indexing on a dask array, which triggers compute_chunk_sizes() and performs an extra full scan before da.percentile runs.
3. _run_dask_cupy_box_plot had the same pattern plus a full map_blocks(cupy.asnumpy) over the dataset before sampling.

Benchmark

head_tail_breaks dask path on a 256×256 gamma-distributed float64 array, chunks=64:

Backend	Metric	Before	After	Ratio	Verdict
dask+numpy	wall_ms (med)	912	339	0.37	IMPROVED

box_plot dask path on 512×512, chunks=128:

Backend	Metric	After	Verdict
dask+numpy	wall_ms (med)	57	OK (no baseline — old path scaled with full-raster scan before percentile)

Test plan

• pytest xrspatial/tests/test_classify.py — 85 tests pass
• Manual smoke: head_tail_breaks dask output has the same bin count as numpy path on the same seed
• Manual smoke: box_plot dask output uses sampled quantiles; verify output classes match numpy path within sampling tolerance

Notes

Sample size for the box_plot dask path is capped at 200,000 elements (or the full dataset if smaller). This matches the pattern used by natural_breaks and keeps the percentile computation O(sample) rather than O(dataset).