Enforce stable_only=True for HTTP/fsspec sources on non-VRT read paths

[brendancol]

Closes #2821

reader.http and reader.fsspec are advanced-tier in SUPPORTED_FEATURES, but stable_only=True was only gated on the VRT dispatch. The eager non-VRT path called read_to_array directly, and the dask path only forwarded stable_only to read_vrt for .vrt sources, so an HTTP/fsspec source read straight through stable_only=True unchecked.

Changes

• Add RemoteStableSourcesOnlyError (a GeoTIFFAmbiguousMetadataError subclass), the non-VRT sibling of VRTStableSourcesOnlyError, and re-export it from xrspatial.geotiff.
• Add _validate_stable_only_remote and call it from open_geotiff (covers the eager/dask/GPU dispatch) and from read_geotiff_dask (direct callers). It rejects http(s):// and fsspec URIs under stable_only=True before any fetch or decode.
• allow_experimental_codecs=True stays the documented unlock; local-file reads still pass.
• Document the gate in docs/source/reference/geotiff.rst.

Backends

The gate lives in open_geotiff ahead of the eager / dask / GPU dispatch, plus a direct check in read_geotiff_dask. CPU eager and CPU dask are covered by the new tests; the GPU dispatch path shares the same open_geotiff gate (no GPU hardware in CI to assert directly).

Test plan

• Eager fsspec source under stable_only=True raises RemoteStableSourcesOnlyError
• Dask fsspec source (via open_geotiff and direct read_geotiff_dask) raises
• http:// source raises before any network fetch
• allow_experimental_codecs=True unlocks the read
• Local-file eager and dask reads still succeed under stable_only=True
• Existing VRT stable_only and release-gate suites still green

[brendancol]

PR Review: Enforce stable_only=True for HTTP/fsspec sources on non-VRT read paths

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

• read_geotiff_gpu (xrspatial/geotiff/_backends/gpu.py:71) takes stable_only but does not gate remote HTTP/fsspec sources on its own. Through open_geotiff the new gate runs ahead of the GPU dispatch, so the public entry point is fine. A direct read_geotiff_gpu(http_url, stable_only=True) call still gets through, the same gap this PR closes for eager/dask. Out of scope for #2821 (eager + dask), but a follow-up would make all three direct entry points match. No GPU in CI to test it here anyway.

Nits (optional improvements)

• The rejection message in _validate_stable_only_remote (xrspatial/geotiff/_validation.py:798) cites release_gate_geotiff.rst and epic #2342, matching _validate_stable_only_vrt. The doc that actually describes this gate is geotiff.rst, not the release-gate checklist. The sibling validator does the same, so leaving it is fine.

What looks good

• The gate sits in open_geotiff ahead of the eager/dask/GPU dispatch (__init__.py:917), with a direct check in read_geotiff_dask (_backends/dask.py), so both the routed and direct dask paths are covered.
• _validate_stable_only_remote mirrors _validate_stable_only_vrt: same early returns, same allow_experimental_codecs unlock, same error family (GeoTIFFAmbiguousMetadataError).
• source is already normalized by _coerce_path before the gate, so path-like inputs are strings by the time _is_fsspec_uri sees them.
• Tests cover eager, dask via open_geotiff, direct read_geotiff_dask, the http:// case (gated before any fetch), the experimental-codecs unlock, and both local-file backends still passing.

Checklist

• Behavior matches the VRT gate it mirrors
• Targeted backends (eager, dask) gated and tested
• Error type/message consistent with the existing VRT path
• Edge cases covered (local file passes, http gated pre-fetch, unlock works)
• No premature materialization (gate runs before any read)
• Benchmark not needed (validation-only change)
• Docstrings present and accurate
• Doc updated (geotiff.rst)

[brendancol]

Review follow-up

No new commits: both findings were resolved without code changes in this PR.

• Suggestion (direct read_geotiff_gpu does not gate remote sources): deferred to #2830. It is GPU-only and was out of scope for the eager + dask finding in #2821. The public open_geotiff(..., gpu=True) path is already covered by the shared gate in this PR; only the direct GPU entry point remains, which #2830 tracks.
• Nit (error message cites release_gate_geotiff.rst rather than geotiff.rst): dismissed. The citation matches the sibling _validate_stable_only_vrt validator on purpose. Changing one and not the other would split the convention, and the review itself flagged leaving it as fine.