zonal: rechunk apply 3D dask output along stacked axis (#2526)

[brendancol]

Summary

• zonal.apply on a 3D dask values array left output chunks at size 1 along the stacked axis because da.stack(layers, axis=2) was not followed by rechunk().
• Rechunk the stacked output back to values_data.chunks[2] in both _apply_dask_numpy and _apply_dask_cupy.
• Adds regression test_apply_dask_3d_axis2_rechunked_2526 and updates the sweep-performance state row for zonal.

Fixes #2526.

Background

Graph probe on a 256x256 raster with chunks=(64, 64, 3) and 3 bands previously yielded:

"apply_dask_3d": {
  "task_count": 176,
  "tasks_per_chunk": 11.0,
  "output_chunks_axis2": [1, 1, 1]
}

The documented anti-pattern from CLAUDE.md is da.stack without a following rechunk — see existing zonal hypsometric_integral audit notes that called out the same shape.

Test plan

• pytest xrspatial/tests/test_zonal.py (126 passed)
• New regression test asserts result.data.chunks[2] == values.data.chunks[2]
• Existing test_apply_3d[dask+numpy] still passes

[brendancol]

Review

Targeted fix, low risk. The da.stack -> rechunk({2: ...}) pattern is the documented remediation for unit-chunk-along-stack-axis (also referenced in the in-tree hypsometric_integral audit).

Verified

• rechunk({2: values_data.chunks[2]}) only touches axis 2; axes 0/1 keep zones_data.chunks.
• meta=np.array(()) / meta=cupy.array(()) from map_blocks propagates through da.stack and rechunk, so dtype/backend metadata is preserved.
• N=1 band edge case: rechunk target becomes (1,), which matches what da.stack already produces - no-op, harmless.
• Both _apply_dask_numpy (line 1691-1694) and _apply_dask_cupy (line 1734-1737) get the fix.
• pytest xrspatial/tests/test_zonal.py -> 126 passed locally.

Notes (not blocking)

• Regression test exercises _apply_dask_numpy only. The cupy path uses the same construction so the fix is symmetric, and this matches the existing test_apply_3d parametrization (['numpy', 'dask+numpy']). cupy parity test can land separately when the cupy CI runner exercises it.
• No benchmark added; existing zonal benchmark already covers apply.

Looks ready to merge.