resample: close test-coverage gaps surfaced by deep-sweep audit

[brendancol]

Summary

Test-only PR closing test-coverage gaps in xrspatial.resample
surfaced by /deep-sweep test-coverage on 2026-05-27. Adds
test_resample_coverage_2026_05_27.py with 70 tests (68 passing,
2 skipped on the documented dask cubic Nx1 limitation tracked in
#2547).

Gaps closed:

• Cat 3 HIGH: Nx1 single-column input across all four backends
  and all eight methods. Existing tests covered 1x1 and 1xN strips
  but never width=1 with multiple rows.
• Cat 2 MEDIUM: NaN parity for cupy and dask+cupy. Existing
  parity tests used random data without NaNs, so the weight-mask
  gate and spline-prepad GPU paths had no NaN coverage.
• Cat 3 MEDIUM: all-equal-value input across methods (zero
  variance / zero gradient).
• Cat 5 MEDIUM: non-default dim names (lat/lon,
  latitude/longitude, channel/lat/lon 3D) propagate
  without being renamed to y/x; per-dim units attrs
  round-trip.
• Cat 3 MEDIUM: empty raster (0 rows / 0 cols) behaviour pin.

Source untouched. Filed #2547 for the cubic-on-dask Nx1 source bug
the audit surfaced; the two skipped tests gate on that fix.

Test plan

• pytest xrspatial/tests/test_resample_coverage_2026_05_27.py
  -> 68 passed, 2 skipped
• pytest xrspatial/tests/test_resample.py xrspatial/tests/test_resample_coverage_2026_05_27.py -> 237
  passed, 2 skipped (no regressions in existing suite)

[brendancol]

PR Review: resample test-coverage closures

Test-only PR, 70 tests across six classes targeting the gaps the deep-sweep audit named. Read the full file in worktree; no source touched.

Blockers

None.

Suggestions

• xrspatial/tests/test_resample_coverage_2026_05_27.py:36-48 — _backend_available() duplicates the runtime check the @cuda_and_cupy_available / @dask_array_available decorators already perform. The duplication is justified here (decorators apply per-class, not per-parameter), but if other _coverage_*.py siblings end up reimplementing the same boolean, a general_checks.backend_available(name) would keep it in one place.
• xrspatial/tests/test_resample_coverage_2026_05_27.py:51-57 — _to_numpy() is generally useful and probably reimplemented in the zonal / polygonize coverage files. Worth a follow-up to promote it to general_checks.py.

Nits

• xrspatial/tests/test_resample_coverage_2026_05_27.py:81 — skip message says "see audit notes" but resample: cubic on dask backends fails for arrays smaller than depth=16 #2547 is the tracking issue. Embedding the issue number in the skip string keeps it greppable once the source bug lands.
• xrspatial/tests/test_resample_coverage_2026_05_27.py:269-278 — the comment claims the (0, 0) cell averages a 2x2 window with one NaN and three 7s to 7. That assumes the average aggregate skips NaN cells. The test passing confirms the behavior; a one-line pointer at _agg_mean / _nan_aware_* would make the assumption explicit instead of implied.
• xrspatial/tests/test_resample_coverage_2026_05_27.py:123,132 — test_nx1_parity_numpy_vs_backends uses .values on the numpy reference and _to_numpy(...) on the others. The helper handles all four backends; using it for the numpy path too removes the visual inconsistency.

What looks good

• Six classes, each with a docstring naming the audit category and the why (not just the what).
• Skip conditions point at the tracking issue (resample: cubic on dask backends fails for arrays smaller than depth=16 #2547) for the dask cubic Nx1 limitation, so the gate is visible in CI output.
• All four backends parametrized where appropriate; CuPy / dask gates use existing decorators.
• Fixed RNG seeds keep NaN parity deterministic.
• TestEmptyRasterRejected accepts (IndexError, ValueError) so a future error-message improvement stays green.
• Filename follows the project's existing _coverage_YYYY_MM_DD pattern (polygonize, rasterize, reproject, zonal already use it).

Checklist

• All implemented backends produce consistent results
• NaN handling exercised across numpy / cupy / dask+cupy
• Edge cases covered (Nx1, all-equal, empty, NaN, non-default dims)
• Dask chunk boundaries — cubic Nx1 limitation tracked separately as resample: cubic on dask backends fails for arrays smaller than depth=16 #2547
• No premature materialization
• Filename and helper imports match project conventions