geotiff: gate non-uniform-coord check on no-georef marker, drop dtype exemption (#2133)

[brendancol]

Summary

• Threads no_georef_marker through the write-metadata validation
  context from the three writer entry points (to_geotiff eager CPU,
  GPU write_geotiff_gpu, dask-VRT streaming).
• _check_write_non_uniform_coords now gates the exemption on the
  marker instead of dtype.kind. A user-authored int-coord grid with
  non-uniform spacing raises NonUniformCoordsError; a marker-stamped
  placeholder grid still skips the check.
• Drops _is_no_georef_sentinel from _coords.py. The predicate was
  test-only after geotiff: to_geotiff silently strips georef on int64 step-1 user coords #2120; the shape check moves into
  test_int_coord_sentinel_2087.py as a local helper.

Backend coverage

numpy, cupy, dask+numpy, dask+cupy. All four write paths pass the
marker into the validator.

Test plan

• pytest xrspatial/geotiff/tests/ (CPU paths) passes locally
• New test_no_georef_attr_migration_2133.py covers the validator
  gap, marker presence on every read backend, marker survival
  across da.copy() / da.assign_attrs(...), and a 3D round-trip
• test_int_coord_sentinel_2087.py::test_non_uniform_int_coords_raise
  still passes (regex relaxed because the validator now catches the
  case before coords_to_transform)
• GPU read-marker tests skip on this box (no CUDA); CI runs them

Closes #2133.

[brendancol]

PR Review: geotiff: gate non-uniform-coord check on no-georef marker, drop dtype exemption (#2133)

Blockers

None.

Suggestions

• docs/source/user_guide/attrs_contract.rst:107-116: the
  _xrspatial_no_georef entry covers round-trip semantics but skips
  the identity-check invariant (is True only). The issue body
  flagged this as worth pinning so a third-party stamp like
  '_xrspatial_no_georef': 'yes' can't flip writer behaviour by
  accident. One line would close that gap.

Nits

• xrspatial/geotiff/_coords.py:41: _has_no_georef_marker(da) lost
  its xr.DataArray annotation when it became tolerant of raw
  numpy / cupy arrays. The docstring covers the new contract, but
  an Any annotation would keep the signature self-documenting.
• xrspatial/geotiff/_validation.py:765-768: the dtype guard now
  accepts both np.floating and np.integer. np.bool_ is a
  subdtype of integer in numpy, so a bool coord array would flow
  through the diff math. Bool coords would be very unusual; mention
  rather than fix.

What looks good

• Marker is threaded through all three writer entry points
  (to_geotiff, _write_vrt_tiled, write_geotiff_gpu) the same way.
• Validator exemption now matches the writer's _has_no_georef_marker
  call, so the two stay in sync.
• _has_no_georef_marker handles raw numpy / cupy arrays passed
  straight to write_geotiff_gpu. The GPU test suite caught the
  earlier AttributeError and the fallback was added before push.
• Sentinel helper moved into the test file as a local predicate.
  Production code no longer carries dead diagnostic code.
• New test file covers the four acceptance criteria from the issue:
  validator gates on marker, marker present on every read backend,
  marker survives copy() / assign_attrs(), 3D round-trip.
• Existing test_non_uniform_int_coords_raise regex relaxed
  correctly. The validator now catches the case before
  coords_to_transform does, so the new message is the authoritative
  one.

Checklist

• Algorithm matches the #2133 acceptance criteria
• All write backends thread the marker consistently
• NaN handling: not applicable for this change
• Edge cases covered by tests (3D, copy, assign_attrs, marker exemption)
• Dask boundaries: not applicable (metadata-only change)
• No premature materialization
• Benchmark not needed for a validator change
• README feature matrix: not applicable (no new public API)
• Docstrings present and updated

[brendancol]

Follow-up: review dispositions

Suggestion (fixed)

• docs/source/user_guide/attrs_contract.rst: added the identity-check invariant note. The entry now spells out that only the exact boolean True flips the writer into no-georef mode, with examples of stamps the writer ignores. Closes the gap from the original review.

Nit (fixed)

• xrspatial/geotiff/_coords.py:_has_no_georef_marker: restored a type annotation. da: Any makes the predicate's tolerance of non-DataArray inputs explicit in the signature, not just the docstring.

Nit (dismissed)

• xrspatial/geotiff/_validation.py:765-768 bool-coord concern: false alarm. np.issubdtype(np.bool_, np.integer) returns False in numpy, so the existing dtype guard already rejects bool coord arrays. No code change needed. The follow-up commit message records the reason.

CPU-side test suites (test_no_georef_attr_migration_2133.py, test_int_coord_sentinel_2087.py, test_no_georef_marker_2120.py, test_ambiguous_metadata_hooks_1987.py) re-run green after the follow-up commit.