Release-gate: windowed-read coords + shifted transform parity (PR 2 of 5 of epic #2341)

[brendancol]

Reason or Problem

Epic #2341 names "windowed reads return unshifted transforms" as a priority risk: a window that returns the file's full-extent transform but a small array is a footgun that downstream spatial functions silently trust. No release-gate test currently pins, for a representative set of files, that a windowed read returns (a) the correct shape, (b) coords that are a slice of the full-file coords, and (c) a transform whose origin matches the window's upper-left, not the file's upper-left.

Proposal

Add xrspatial/geotiff/tests/test_release_gate_windowed_reads_2341.py.

For each file in the representative corpus, for both eager (open_geotiff(..., window=...)) and dask (read_geotiff_dask(..., window=...)) read paths:

• Open the file with no window; record the full transform T_full and full coords.
• Open the file with a window strictly interior to the raster (e.g. Window(col_off=64, row_off=32, width=128, height=64)).
• Assert the returned shape equals the window's (height, width).
• Assert result.coords['x'] equals full.coords['x'][64:64+128] (and similar for y).
• Assert result.attrs['transform'] equals T_full * Affine.translation(window.col_off, window.row_off) exactly — no float drift.
• Assert the seven release-attr keys other than transform are unchanged.

Assertions inlined per-file (see PR 1 of 5 of epic #2341 for the pattern). Add a row to docs/source/reference/release_gate_geotiff.rst under "Local GeoTIFF read and write" citing the new file.

Stakeholders and Impacts

Touches one new test file. No production code change unless the test surfaces a real bug, in which case the bug fix is in scope for this PR.

Out of scope: overview/sidecar (PR 3), codec round-trip (PR 4), negative tests (PR 5).

Drawbacks

Some overlap with existing window tests under xrspatial/geotiff/tests/ (e.g. test_gpu_window_band_1605.py, test_http_stripped_window_max_pixels_issue_A_1842.py), but those focus on specific subsystems. This file is the single citation for the release-gate row.

Alternatives

• Use rasterio's window-math as the oracle. Rejected for the transform assertion: we want to lock the contract that the xrspatial transform exactly equals T_full * translation(col_off, row_off), which is the spec, not whatever rasterio happens to do.

Additional Notes or Context

Parent epic: #2341. PR 2 of 5.