test(classify): cover all-NaN degenerate input across classifiers

[brendancol]

Adds all-NaN regression tests for the classifiers that lacked them.

std_mean and maximum_breaks already return an all-NaN result on
all-NaN input, so they get plain passing tests. head_tail_breaks,
percentiles, and box_plot currently crash on all-NaN input, so their
tests are xfail (non-strict) and reference the source bug in #3510. Flip
them to plain assertions once #3510 is fixed.

equal_interval and natural_breaks already had all-NaN coverage.

Test-only change, no source modified. Ran locally on a CUDA host:
4 passed, 3 xfailed.

Found by the test-coverage sweep. Fixes the coverage half of #3510.

[brendancol]

Review: degenerate-input regression tests for classify

Blockers: none.

Suggestion (addressed in a8cbf28): the all-NaN tests ran on numpy only, but the degenerate path diverges by backend. Running all five on dask:

• std_mean and maximum_breaks return all-NaN on both eager and dask
• head_tail_breaks crashes eager, but its dask path returns all-NaN cleanly because of a total_count == 0 guard the eager path lacks
• percentiles and box_plot crash on dask too

Added five @dask_array_available all-NaN tests: passing for std_mean, maximum_breaks, and head_tail_breaks (dask); xfail(#3510) for percentiles and box_plot (dask). Also wrapped the std_mean all-NaN tests in warnings.catch_warnings() to drop three empty-slice RuntimeWarnings.

Note for #3510: the head_tail_breaks crash is eager-only (numpy/cupy). The source fix should bring the eager paths in line with the existing dask guard.

What looks good: xfail(strict=False) is the right call with parallel sweeps in flight, since the tests assert the real all-NaN result and the xfails flip to pass exactly when #3510 lands. Full test_classify.py: 107 passed, 5 xfailed.