Vectorize _sort_and_stride 3D reindex, drop deepcopy (#3381)

[brendancol]

Closes #3381

What changed

• The 3D branch of _sort_and_stride deep-copied the entire values array and then reindexed each layer by sorted_indices in a Python loop. It now does the reindex with one vectorized fancy-index, values.reshape(n, -1)[:, sorted_indices]. Fancy indexing returns a fresh array, so the input is never mutated and the explicit copy is gone.
• Removed the now-unused import copy.

This path runs in crosstab() on the numpy backend and once per block on dask+numpy, so the redundant full-array copy was paid per block at scale. The cupy backend rejects 3D input, so it is unaffected.

Backend coverage

numpy, dask+numpy (the only backends that reach the 3D _sort_and_stride path).

Test plan

• New test_sort_and_stride_3d_no_mutation_and_parity pins input non-mutation and layer-wise sort parity.
• Existing 3D crosstab tests pass (test_crosstab_3d_count, test_crosstab_3d_agg_method, test_nodata_values_crosstab_3d), all of which already assert the input is unmodified.
• Full test_zonal.py (194) and test_zonal_backend_coverage_2026_05_27.py (43) pass.

[brendancol]

PR Review: Vectorize _sort_and_stride 3D reindex, drop deepcopy (#3381)

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• test_zonal.py (new test_sort_and_stride_3d_no_mutation_and_parity): the parity assertion recomputes expected with the same expression the implementation now uses (values.reshape(n, -1)[:, sorted_indices]), so it is not an independent oracle. The non-mutation assertion also would not have failed against the old deepcopy code, since the old path mutated a copy rather than the input. The test still earns its place as a guard against a future in-place-mutation or shape regression, and test_crosstab_3d_count / test_crosstab_3d_agg_method already provide the independent reference-value check. No change needed, just noting the test's scope.

What looks good

• The replacement is equivalent: reshape returns a view when the array is contiguous, and the fancy-index [:, sorted_indices] always allocates a fresh array, so the input is never touched (confirmed empirically).
• The deepcopy and the Python row loop both collapse into one vectorized op.
• Removing the now-unused import copy is the right cleanup.
• Backend scope is correct: only numpy and dask+numpy reach the 3D branch; cupy rejects 3D upstream, so nothing GPU-side is affected.
• Existing 3D crosstab tests already assert the input is unmodified and check against reference values.

Checklist

• Algorithm matches reference: n/a (mechanical refactor, output identical)
• All implemented backends consistent: yes (numpy + dask+numpy share the helper; cupy unaffected)
• NaN handling correct: unchanged (NaN filtering happens downstream of this helper)
• Edge cases covered: yes (existing 3D tests + new unit test)
• Dask chunk boundaries handled: n/a (per-block helper, no chunk logic changed)
• No premature materialization or unnecessary copies: improved (one full-array copy removed)
• Benchmark exists or not needed: existing benchmarks/benchmarks/zonal.py covers stats and regions; crosstab 3D is unbenchmarked but this is a refactor, not a new function
• Docstrings present and accurate: n/a (private helper)