zonal: make crop() return (0, 0) on all backends when no zone matches

[brendancol]

Summary

• Fix backend divergence in xrspatial.zonal.crop(). When the requested zone_ids are not present in the zones raster, numpy/cupy used to return shape (0, 0) while dask silently returned the full input raster unchanged. Now all four backends return (0, 0).
• Reworked _crop (CPU) and _crop_bounds_dask to return an explicit found flag (1/0) so the call site picks the empty slice deterministically instead of relying on inverted bounds.
• Documented the no-match behaviour in the crop() docstring Notes section.

Backend coverage: numpy, cupy, dask+numpy, dask+cupy (parametrized regression tests).

Closes #2561

Test plan

• pytest xrspatial/tests/test_zonal.py -k crop, 20/20 pass (existing 11 + 9 new parametrized cases)
• pytest xrspatial/tests/test_zonal.py, full file 148/148 pass
• Manual repro from the issue body: numpy and dask both return (0, 0) for absent zones
• CI (numpy, cupy where available, dask+numpy, dask+cupy)

[brendancol]

PR Review: zonal: make crop() return (0, 0) on all backends when no zone matches

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

• xrspatial/tests/test_zonal.py:1959 — test_crop_partial_zone_ids_2561 only checks that partial matches behave the same as the "present-only" subset. It does not pin an expected shape, so if both code paths drifted together the test would still pass. Adding assert result_partial.shape == (2, 2) (the bounding box of zone_id=1) makes the regression bar explicit.

Nits (optional improvements)

• xrspatial/zonal.py:2599 — the final return uses a literal 1 for the found flag while the docstring describes it as the variable found. Using return top, bottom, left, right, found is a touch more self-documenting; the local found is already 1 by the time we reach this line. Cosmetic only.
• xrspatial/tests/test_zonal.py:1916 — _crop_backends_2561() runs at module import time inside @pytest.mark.parametrize(...). Other tests in this file build backend lists the same way, so this is consistent. Flagging it only as a heads-up for future readers.

What looks good

• The found flag convention is symmetric across _crop and _crop_bounds_dask, so the call site is uniform.
• Slicing with values[0:0, 0:0] preserves underlying array type (numpy / cupy / dask) and dim names. Verified by manual repro; no special-casing per backend needed.
• The numpy branch short-circuits with if found == 0: return 0, 0, 0, 0, 0 right after the top-pass, so it does not waste work scanning for bottom / left / right.
• The crop() docstring Notes section spells out the new (0, 0) contract for all four backends.
• Cross-backend tests use create_test_raster and gate on has_dask_array / has_cuda_and_cupy, matching the existing pattern in this file.
• Existing test_crop_nothing_to_crop (where zone_id=0 is actually present in the raster) still passes; the behavior change only affects the genuinely-absent case.

Checklist

• Algorithm matches reference/paper — N/A, this is a behaviour contract change
• All implemented backends produce consistent results — (0, 0) everywhere
• NaN handling is correct — unchanged from prior behaviour
• Edge cases are covered by tests — absent, partial, happy path
• Dask chunk boundaries handled correctly — uses the existing reduction approach
• No premature materialization or unnecessary copies — dask.compute was already there
• Benchmark exists or is not needed — bug fix, no perf impact
• README feature matrix updated (if applicable) — no API surface change
• Docstrings present and accurate — Notes section updated

[brendancol]

Follow-up review (after ca1cae2)

Re-reviewed the diff after the follow-up commit. Both addressable findings from the first pass are resolved.

Disposition

• Suggestion: pin shape in test_crop_partial_zone_ids_2561 — fixed. assert result_partial.shape == (2, 2) now anchors the bounding box of zone_id=1, so the test can no longer pass by accidental co-drift of both code paths. Tests still green (20/20).
• Nit: _crop final return uses found instead of literal 1 — fixed. Self-consistent with the docstring now.
• Nit: _crop_backends_2561() called at module import — dismissed. Matches the existing convention in test_zonal.py and elsewhere; changing it would diverge from the project pattern.

No new findings on this pass. Ready for CI.