crosstab: validate backend compatibility on the 3D path

[brendancol]

Closes #2640

crosstab() only validated backend/shape compatibility on the 2D path. A 3D call with mismatched backends (numpy zones plus a dask 3D values, for example) sailed past validation and crashed deep inside the dask/numba machinery with TypeError: 'NoneType' object is not subscriptable, which gives the caller no hint that the real problem is a backend mismatch.

This adds an early backend check to the 3D path. validate_arrays() can't be reused here because it requires equal shapes and zones is 2D while values is 3D, so the check compares _classify_backend() of each input and raises a clear ValueError when they differ.

• Adds a backend-compatibility check for 3D crosstab() inputs, raising ValueError up front instead of crashing later.
• New test covers both directions of mismatch (numpy zones + dask values, and the reverse).

Backend coverage: the validation runs for all backends. Matching-backend 3D calls (numpy/numpy, dask/dask) are unchanged.

Test plan:

• pytest xrspatial/tests/test_zonal.py -k crosstab_3d
• pytest xrspatial/tests/test_zonal.py (full module, no regressions)

[brendancol]

PR Review: crosstab: validate backend compatibility on the 3D path

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• xrspatial/zonal.py:1416-1430 -- the new check covers the backend mismatch but not the "early shape" half of the original report. Spatial shape is still validated later (zones.shape != values.shape[1:]), which only fires after layer is resolved and the array is transposed. That's fine in practice: a real shape mismatch still raises a clear "Incompatible shapes" ValueError (verified). The one rough edge is that if layer is also invalid, the user sees "Invalid layer" first. Not worth reordering for this PR; flagging it so the call is on record.
• xrspatial/tests/test_zonal.py:1354 -- the test is gated on dask being available, which is right since both cases need a dask array. There's no numpy-vs-cupy mixed case, but that needs a GPU runner, so leaving it out is reasonable.

What looks good

• The new check sits right next to the existing 2D validate_arrays() call, so the two paths read symmetrically.
• Reusing _classify_backend() keeps the backend names consistent with the rest of the codebase (numpy / cupy / dask+numpy / dask+cupy).
• The comment explains why validate_arrays() can't be reused here (it needs equal shapes), which is the non-obvious part.
• The test checks both directions of the mismatch and asserts the error message, not just the exception type.
• Matching-backend 3D calls (numpy/numpy, dask/dask) are unchanged; the full test_zonal.py passes (170 tests).

Checklist

• Algorithm matches reference/paper: n/a (validation hardening, no algorithm change)
• All implemented backends produce consistent results: yes, the check is backend-agnostic
• NaN handling correct: unchanged
• Edge cases covered by tests: mixed-backend, both directions
• Dask chunk boundaries handled correctly: unchanged
• No premature materialization: yes, the check only inspects backend type, no compute
• Benchmark exists or not needed: not needed (validation path)
• README feature matrix updated: n/a (no new function, no backend support change)
• Docstrings present and accurate: unchanged public signature

[brendancol]

PR Review (follow-up after merge with main)

Re-reviewed after merging origin/main (which brought in #2619's 3D crosstab cupy/dask+cupy test coverage). No code changes were needed in response to the first review; the two nits were both dismissed with reasons:

• Early-shape ordering nit: dismissed. Same-backend spatial mismatches already raise a clear "Incompatible shapes" ValueError. The "Invalid layer" ordering edge case is pre-existing and reordering it would expand this PR's scope past the backend gap it targets.
• No numpy-vs-cupy mixed test: dismissed. That case needs a GPU runner not present in CI.

The merge was conflict-free. The backend check survived intact and the full zonal suite passes (218 tests across test_zonal.py and the backend-coverage file, including the new #2619 3D cupy cases). No new findings.