Decouple GPU alias test from real CUDA runtime (#2515)

[brendancol]

Closes #2515.

Summary

• test_gpu_alias_accepts_old_values_without_validation_error_1560 was failing on hosts where CuPy imports but the CUDA runtime can't run. The test sends gpu='strict' through read_geotiff_gpu to confirm the deprecated alias passes value validation. On a broken-CUDA host the call hits _preflight_cuda_runtime (the documented behaviour from issue geotiff: preflight CUDA in read_geotiff_gpu so broken-driver errors are clean #1903) and the resulting RuntimeError is not in the test's accepted-exception tuple.
• Fix: reuse the existing _install_cupy_stub_1903 helper to stub cupy with a preflight that reports one device. The alias call then reaches the file-read stage on every host and fails with FileNotFoundError, which is what the alias check needs to assert.
• The preflight RuntimeError itself is unchanged; it's the correct contract for geotiff: preflight CUDA in read_geotiff_gpu so broken-driver errors are clean #1903.

Backend coverage

Test-only change. No public API surface, no backend dispatch, no GPU code paths touched. Alias-validation coverage now runs uniformly on CPU-only, broken-CUDA, and GPU hosts.

Test plan

• pytest xrspatial/geotiff/tests/gpu/test_kernels_and_kwargs.py::test_gpu_alias_accepts_old_values_without_validation_error_1560 passes on a CuPy-installed-but-CUDA-unusable host.
• pytest -q xrspatial/geotiff/tests -m 'not slow' is fully green on that same host (5913 passed, 0 failed).
• CI validates the same fix on CPU-only and (where available) GPU runners.

[brendancol]

PR Review: Decouple GPU alias test from real CUDA runtime (#2515)

Blockers

None.

Suggestions

None.

Nits

• xrspatial/geotiff/tests/gpu/test_kernels_and_kwargs.py:76-81 — the _install_cupy_stub_1903 docstring says "Used on machines without cupy installed; lets us exercise the preflight failure path on CPU-only CI." This PR also uses it on broken-CUDA hosts for the preflight success path, which is a wider role than the docstring describes. A one-line tweak to the helper's docstring would keep its contract aligned with its callers.

What looks good

• The fix targets the test, not the preflight contract. The PR body says so directly.
• Reuses _install_cupy_stub_1903 instead of inventing a new pattern.
• monkeypatch.setitem(sys.modules, ...) is function-scoped and auto-restored, so sibling tests aren't affected.
• Drops ImportError from the accepted-exception tuple now that the stub guarantees cupy resolves. An unexpected ImportError will now fail loudly rather than be silently tolerated.
• Test-only. No public API, no GPU code paths, no backend dispatch touched.
• Verified locally: 5913 passed, 0 failed on the CuPy-installed-but-CUDA-unusable host that reproduced the original bug.

Checklist

• Algorithm matches reference/paper — N/A (test-only)
• All implemented backends produce consistent results — N/A (alias coverage now uniform across hosts)
• NaN handling is correct — N/A
• Edge cases covered — alias validation passes on CPU-only, broken-CUDA, and GPU hosts
• Dask chunk boundaries handled correctly — N/A
• No premature materialization or unnecessary copies — N/A
• Benchmark exists or is not needed — not needed (test-only)
• README feature matrix updated — N/A
• Docstrings present and accurate — yes

[brendancol]

PR Review: Decouple GPU alias test from real CUDA runtime (#2515) — follow-up pass

Blockers

None.

Suggestions

None.

Nits

None. The earlier nit on helper docstring scope was handled in 3164b23; the docstring now covers both the failure-path and success-path callers.

What looks good

• The docstring update names the alias-validation use case directly (mentions #2515), so anyone reading later can see why the success-path mode exists.
• Diff is still scoped to one test file. No drift into unrelated code.
• All 163 tests in the file pass after the docstring tweak.

Checklist

• Algorithm matches reference/paper — N/A (test-only)
• All implemented backends produce consistent results — N/A
• NaN handling is correct — N/A
• Edge cases covered — alias validation runs uniformly across hosts
• Dask chunk boundaries — N/A
• No premature materialization — N/A
• Benchmark — not needed
• README feature matrix — N/A
• Docstrings present and accurate — yes (helper docstring is broader after the follow-up commit)