Add kriging test coverage (#2921)

[brendancol]

Closes #2921

Test-only PR. No source changes. Fills coverage gaps the test-coverage sweep found in xrspatial/interpolate/_kriging.py.

What changed

• test_dask_return_variance_matches_numpy: the dask+numpy return_variance=True branch (_kriging_dask_numpy._chunk_var) had no test. The dask+cupy variance path was already covered; this matches that pattern (atol=1e-12, since the variance values sit near 1e-14).
• nlags coverage: added a non-default nlags=5 case plus the invalid-input paths (nlags=0 and nlags=-1 raise ValueError, nlags=2.5 raises TypeError).
• test_two_point_warns_few_lag_bins: a two-point input produces fewer than 3 lag bins and warns; asserts the UserWarning and finite output.
• test_kriging_all_nan_points: all-NaN input raises ValueError (only the idw equivalent was tested before).
• test_output_metadata: output preserves template coords, dims, attrs, and name.
• test_single_point: xfail(strict, raises=ValueError) documenting the single-point crash. That is a real source bug (kriging() crashes on single-point input with "zero-size array to reduction maximum" #2920); the fix is left to that issue so this PR stays test-only.

Backend coverage

numpy, cupy, dask+numpy, dask+cupy all have parity tests. The new dask+numpy variance test closes the one backend branch that had no coverage. All GPU tests ran green on a CUDA host.

Test plan

• pytest xrspatial/tests/test_interpolation.py -k "Kriging or kriging" -> 22 passed, 1 xfailed
• cupy / dask+cupy tests executed (not skipped) on a CUDA host

[brendancol]

Review: Add kriging test coverage (#2928)

Test-only PR. I read the full test file and the kriging source, and ran the suite on a CUDA host (49 passed, 1 xfailed).

Blockers

None.

Suggestions

None.

Nits

• test_output_metadata sets template.attrs['res'] on the local template. That is fine because _make_template builds a fresh DataArray per call, so there is no leakage into other tests. No change needed; noting it for the record.

What looks good

• The dask+numpy return_variance branch (_kriging_dask_numpy._chunk_var) was the one backend path with no test. The new test_dask_return_variance_matches_numpy closes it, and uses atol=1e-12 to match the existing dask+cupy variance test (the variance values sit around 1e-14, where rtol is meaningless).
• The single-point xfail is strict=True with raises=ValueError, so it converts to a failure the moment #2920 is fixed. The source fix is correctly kept out of this PR.
• Invalid-nlags paths are split by exception type (ValueError for below-min, TypeError for non-int), matching _validate_scalar.
• All four backends now have parity tests; cupy and dask+cupy ran (not skipped) on the CUDA host.

Checklist

• No source changes; behavior unchanged
• All implemented backends exercised, GPU validated
• NaN / edge cases covered (all-NaN input, two-point sparse, single-point xfail)
• Dask variance branch now tested
• No premature materialization introduced
• Benchmark not applicable (test-only)
• Docstrings / README not applicable (no API change)