Document num_sample=None in natural_breaks docstring

[brendancol]

Closes #3501.

natural_breaks already accepts num_sample=None (use all data) the same
way quantile, maximum_breaks, and percentiles do, but its docstring
only described num_sample as int and never mentioned None. This brings
the wording in line with the other three classifiers.

Docstring only. No signature or behavior change.

Verified natural_breaks(agg, k=5, num_sample=None) on both the numpy and
cupy backends. pytest xrspatial/tests/test_classify.py -k natural_breaks
passes (22 passed).

[brendancol]

Review: document num_sample=None in natural_breaks

Read the full natural_breaks / _compute_natural_break_bins / _run_dask_natural_break paths, not just the diff. Documenting that None is accepted is the right call and matches quantile, maximum_breaks, and percentiles. One wording issue.

Suggestion

xrspatial/classify.py:851-852 -- the new parenthetical "(safe for numpy/cupy, automatically capped for dask)" reads like it was copied from quantile/percentiles, where None just runs an exact percentile over the finite values. That clause doesn't describe natural_breaks:

• On numpy/cupy, None makes _compute_natural_break_bins fit Jenks on the whole array. That's the O(n^2) path this same docstring warns about two lines up ("When n is large, we should fit the model on a small sub-sample").
• _compute_natural_break_bins (line 684) raises MemoryError when the two Jenks matrices would exceed half of available RAM. So None on a large raster is the opposite of "safe" -- it's the case most likely to blow up.
• For dask it isn't capped to a small sample either. _run_dask_natural_break sets num_sample = num_data and fits Jenks on all the data via indexed access.

Suggest rewording so it says None fits on all data, and that this is the expensive O(n^2) case guarded by a memory check, instead of calling it "safe".

What looks good

• Documents the None option, in line with the other three samplers.
• Docstring-only, non-breaking. num_sample=None verified on numpy and cupy.

Checklist

• No code change; behavior unchanged
• Docstring accurately describes None for the Jenks case (see suggestion)
• API consistency improved
• No tests needed (docstring only)