Fix to_geotiff docstring tier for extra_tags pass-through

[brendancol]

Closes #3593

• Moves "extra_tags pass-through" from the [advanced] bullet of to_geotiff's release-contract tier list to [experimental], matching SUPPORTED_FEATURES['writer.extra_tags'] and the _validate_write_rich_tag_optin gate (which raises ValueError on a fresh DataArray with attrs['extra_tags'] unless allow_experimental_codecs=True). The bullet now names the opt-in and the round-trip exemption, and mentions gdal_metadata_xml which shares the gate.
• Clarifies the photometric parameter doc: the override rides attrs['extra_tags'] (not a kwarg) and sits at the experimental tier.
• Adds test_to_geotiff_docstring_extra_tags_tier_matches_supported_features to tests/unit/test_signatures.py (Section 3, next to the existing rich-tag gate tests) so the docstring cannot drift from the tier map again.
• Updates the api-consistency sweep state CSV row for geotiff.

Doc-only change, no behavior touched, so backend coverage is unaffected (the gate itself already fires identically on the CPU and GPU write paths).

Test plan:

• pytest xrspatial/geotiff/tests/unit/test_signatures.py (197 passed, includes the new test)
• pytest xrspatial/geotiff/tests/unit/test_photometric.py (37 passed)
• flake8 clean on both touched source files
• Empirical check on a CUDA host: fresh attrs['extra_tags'] write raises without the flag, succeeds with it, and the cupy write round-trip still works

[brendancol]

Doc-only change; I checked the claims against the code rather than just the prose.

• The new [experimental] bullet matches _validate_write_rich_tag_optin: fresh attrs['extra_tags'] / attrs['gdal_metadata_xml'] raise without allow_experimental_codecs=True, and the _xrspatial_geotiff_contract marker exempts round-tripped attrs. SUPPORTED_FEATURES carries both keys at 'experimental', so the docstring and tier map now agree.
• The drift-guard test scopes its string checks to the region before "Parameters", which keeps the photometric parameter doc (which legitimately mentions the extra_tags override) out of the assertion. The bullet slicing on '* [' holds because [internal-only] follows the experimental bullet.

Nits, not blocking:

• The gate also fires on a dict attrs['gdal_metadata'] (#3320); the revised bullet doesn't name that third attr. The runtime error does, so this is fine to leave.
• The test pins the two SUPPORTED_FEATURES tiers directly, which overlaps with the release-gate suite. Harmless redundancy that makes the test self-contained.

No blockers.