Gate remote sources in read_geotiff_gpu under stable_only (#2867)

[brendancol]

Closes #2867

read_geotiff_gpu(..., stable_only=True) accepted the kwarg but never enforced it. For HTTP / fsspec sources the GPU path routes through the CPU fallback, so a direct GPU call could read an advanced-tier remote source that open_geotiff and read_geotiff_dask reject. The GPU docstring also called stable_only a no-op, which was wrong for remote sources.

• Call _validate_stable_only_remote(...) in read_geotiff_gpu before the cupy import, the CUDA preflight, and the remote CPU fallback, the same way open_geotiff and read_geotiff_dask call it. Running it ahead of the import means the rejection is independent of GPU availability.
• Fix the GPU docstring so stable_only is described as an enforced gate, not a no-op.

Backend coverage: the gate lives in the shared read_geotiff_gpu entry point, which serves the eager GPU (numpy/cupy) and chunked GPU (dask+cupy) reads. The other backends already had the gate.

Test plan:

• Added GPU coverage to test_stable_only_remote_2821.py: fsspec and http sources rejected under stable_only=True, plus an experimental-optin unlock test gated behind cupy + CUDA.
• Rejection tests pass on a CPU-only host (the gate fires before the cupy import).
• pytest xrspatial/geotiff/tests/test_stable_only_remote_2821.py passes (12 passed locally with a GPU present).

[brendancol]

PR Review: Gate remote sources in read_geotiff_gpu under stable_only (#2867)

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• The unlock test (test_gpu_direct_fsspec_source_allowed_with_experimental_optin) re-declares the _gpu_available / _HAS_GPU helper that a few other geotiff test modules already define locally (read/test_basic.py, unit/test_predictor.py). There is no shared fixture for it, so the local copy matches the existing pattern in this package. Fine to leave as is.

What looks good

• The gate runs before the cupy import, the CUDA preflight, and the remote CPU fallback, so the rejection does not depend on GPU availability. That is what lets the two rejection tests run on CPU-only CI, which is the right call.
• The call signature matches how open_geotiff and read_geotiff_dask invoke _validate_stable_only_remote (same kwargs, same local-import style).
• The gate sits ahead of the chunks is not None branch, so it covers both the eager and chunked GPU paths, not just one.
• The docstring no longer claims stable_only is a no-op. The wording now matches actual behavior and the geotiff.rst note that already said all three paths run the check.
• Tests cover fsspec and http rejection plus the experimental-codecs unlock, mirroring the existing eager/dask cases.

Checklist

• Behavior matches the sibling readers (open_geotiff, read_geotiff_dask)
• Gate fires before any GPU/network work
• Covers eager and chunked GPU paths
• Edge cases covered by tests (fsspec, http, unlock)
• No premature materialization or unnecessary copies
• Benchmark not needed (validation-only change)
• README feature matrix not affected (no new function, no backend change)
• Docstring present and accurate