Refactor GeoTIFF Phase 2: centralize transform/georef contract

[brendancol]

Closes #2225
Part of #2211

Summary

Centralises the transform/georef contract behind a single resolver so every backend (eager, GPU, dask, VRT, read and write) makes the same decision for matching inputs.

• Adds GeorefResolution dataclass and resolve_georef() in _coords.py. Fields: transform, georef_status, dropped_reason, applied_no_georef_marker. The resolver owns coord-to-transform inference, the no-georef marker, degenerate-axis policy, and rotated-affine drop policy.
• Routes _attrs._compute_georef_status through the resolver so the read-side bucket decision is computed in one place.
• Replaces the inline _transform_from_attr -> _coords_to_transform -> _require_transform_for_georeferenced ladder in _writers/eager.py (two sites: to_geotiff and _write_vrt_tiled) and _writers/gpu.py with one resolve_georef(data) call.
• The synthesised GeoInfo from _backends/vrt.py::_vrt_to_synthetic_geo_info already flowed through _compute_georef_status, which now routes through the resolver. Added a docstring note.

The public API is unchanged. georef_status values, attrs['transform'] precedence, and the fail-closed behaviour on degenerate coords (#1945) all match prior behaviour.

Tests

New file xrspatial/geotiff/tests/test_georef_resolver_parity_2211.py (22 tests) covers the six fixture cases the issue calls out plus cross-site parity:

• y/x, lat/lon, latitude/longitude dim names
• row/col (no-georef placeholder + _NO_GEOREF_KEY marker)
• transform-only via attrs['transform']
• CRS-only via attrs['crs']
• rotated-affine GeoInfo -> rotated_dropped bucket
• writer-and-reader-agree assertions on full and no-georef cases

Verification

• pytest xrspatial/geotiff/tests/ -> 4,729 passed, 45 skipped, 1 xfailed (pre-existing lz4 failure is unrelated and reproduces on main).
• test_georef_status_2136.py and test_allow_rotated_* suites: all green.

Notes

• xrspatial/geotiff/_writers/vrt.py is the public write_vrt mosaic creator. It does not handle DataArrays or build transforms, so the resolver does not apply there. The per-tile VRT write path lives in _writers/eager.py::_write_vrt_tiled and was updated.
• Kept _compute_georef_status_from_parts (VRT inline shim) as-is so VRT branches do not have to build a fake GeoInfo to feed the resolver; its decision table mirrors the resolver's reader-path branch.

[brendancol]

PR Review: Refactor GeoTIFF Phase 2: centralize transform/georef contract

Blockers (must fix before merge)

• None. Behavior is preserved and the full geotiff suite (4,729 tests) passes.

Suggestions (should fix, not blocking)

• xrspatial/geotiff/_coords.py:723-728 -- The writer path bucketed any "transform present + no CRS" input as coords, which masked the distinction between transform_only (transform came from attrs['transform']) and coords (transform was derived from coord arrays). The resolver now tracks the source and picks the right bucket. Addressed in a follow-up commit on this PR.
• xrspatial/geotiff/_coords.py:708-720 -- The writer path's CRS lookup ignored string-valued attrs['crs']. Some pipelines stash a WKT string there, and attrs_to_metadata already folds it into crs_wkt. The resolver was missing the same fold, so a CRS-only DataArray with a string crs attr would have been classified as none. Same follow-up commit fixed it.

Nits (optional improvements)

• xrspatial/geotiff/_coords.py:411-489 -- The bare-input branch (no DataArray, no GeoInfo) is convenient for tests but no production call site uses it today. Worth a docstring line noting that it is a fixture-only entry point.
• xrspatial/geotiff/_attrs.py:885 -- The thin _compute_georef_status delegation could be inlined once all callers have moved to resolve_georef. Leaving the shim is fine for this PR; the _* name stays stable for the other in-flight PRs in #2211.

What looks good

• GeorefResolution is a frozen dataclass with Literal-typed georef_status, so callers get type narrowing for free.
• The dropped_reason strings (rotated_affine_dropped, no_georef_marker, no_transform_inferred, no_inputs) make failures easy to read without diving into the resolver source.
• 22 new parametrized tests cover the six fixture cases the issue called out plus cross-site parity assertions.
• The unused _transform_from_attr / _coords_to_transform imports in the two writer modules are gone, so resolve_georef is the only path.

Checklist

• Decision table mirrors _compute_georef_status with the writer/reader split documented.
• Backend parity: eager numpy, dask+numpy, GPU eager, GPU dask, and VRT (eager + chunked) all route through the resolver.
• No change to nodata or masking paths.
• Edge cases (1x1, 1xN, degenerate axis, _NO_GEOREF_KEY, rotated affine) all exercised.
• No chunk-boundary code touched.
• Writer paths still read attrs only; nothing materialises.
• [n/a] Benchmarks: pure refactor.
• [n/a] README feature matrix: no public API change.
• Docstrings on GeorefResolution, resolve_georef, and the rerouted _compute_georef_status.

[brendancol]

PR Review (follow-up): Refactor GeoTIFF Phase 2

Re-reviewed after commit 3f0eddc.

Blockers

• None.

Suggestions

• None outstanding. The two suggestions from the first review (writer-side bucket split and string-attrs['crs'] fold) are addressed.

Nits (left for follow-up, with reasons)

• xrspatial/geotiff/_coords.py -- The bare-input branch (no DataArray, no GeoInfo) is still fixture-only. Not addressed here because adding a "fixture-only" disclaimer to a public-looking helper is a doc-only change that competes for diff with the other in-flight PRs in #2211. Worth a follow-up.
• xrspatial/geotiff/_attrs.py -- The thin _compute_georef_status shim could be inlined once #2226 / #2227 / #2229 land. Keeping the name now avoids rebase friction with those siblings.

What still looks good

• transform_source is tracked locally and only sets the bucket; no side-effects.
• The string-crs -> crs_wkt fold matches the attrs_to_metadata precedence exactly, so a CRS-only DataArray with a string attrs['crs'] no longer mis-classifies as none.
• Parity test was updated to assert the new transform_only bucket on a transform-only DataArray.
• All 77 tests in the focused suite (parity, georef_status, allow_rotated) pass; the broader 4,729-test geotiff suite still passes too.

Disposition

Cleared for the next stage of #2211. No remaining actionable findings in this PR.