Positive VRT mosaic tests over GeoTIFF sources (#2369)

[brendancol]

Closes #2369. Sub-task of #2342.

Positive coverage for the simple VRT mosaic subset that the release contract lists as supported. The other VRT tests touch this subset while exercising codecs, dtypes, or chunk policies; nothing pinned the contract by itself.

Six tests, all driving the public read_vrt:

• eager 2x1 horizontal mosaic
• eager 2x2 mosaic with one distinct value per tile (catches placement swaps loudly)
• windowed read aligned with the source seam
• dask read of both mosaics with multiple chunks
• multi-band 2x1 mosaic

Each test asserts values, coords, and key attrs (crs, transform, nodata).

Backend coverage

• numpy: yes
• dask + numpy: yes
• cupy / dask + cupy: not in scope (the release contract only promises eager + dask for VRT)

Test plan

• pytest xrspatial/geotiff/tests/test_vrt_simple_mosaic_2369.py passes locally
• CI runs green on the same file

[brendancol]

PR Review: Positive VRT mosaic tests over GeoTIFF sources (#2369)

Blockers

None.

Suggestions

• xrspatial/geotiff/tests/test_vrt_simple_mosaic_2369.py:118,146,184: the three fixtures use tempfile.mkdtemp(prefix='tmp_2369_...'), which leaks directories between runs. The implementation notes called for tmp_path, and most newer tests in this directory use the pytest fixture. Switching gets auto-cleanup and keeps the prefix-based collision protection.
• xrspatial/geotiff/tests/test_vrt_simple_mosaic_2369.py:240-254: _assert_attrs_ok checks the pixel size on the transform but not origin_x / origin_y. A transform with the wrong origin would still pass. Asserting transform[2] and transform[5] against the fixture origin would catch that case.

Nits

• xrspatial/geotiff/tests/test_vrt_simple_mosaic_2369.py:333: h, w_total = expected.shape -- w_total is unused. Drop it or reference it in the comment.
• xrspatial/geotiff/tests/test_vrt_simple_mosaic_2369.py:410-431: no dask variant of the multi-band test. The eager case covers the contract; a chunked multi-band read would also pin down per-block band handling.

What looks good

• Six tests, narrow scope, exactly the supported subset from the issue.
• 2x2 fixture uses one distinct constant per tile so placement bugs show up in the value diff instead of as numeric drift.
• Each test asserts values, coords, and key attrs (crs, transform, nodata).
• Tile builders are short and shared cleanly across fixtures.

Checklist

• Algorithm matches reference: synthesises mosaics from to_geotiff + internal write_vrt, mirroring test_vrt_lazy_chunks_1814.py.
• All implemented backends produce consistent results: dask compute compared against eager.
• NaN handling: not exercised in this PR; mosaics use finite data plus declared nodata sentinel.
• Edge cases: windowed read on a 2x1 seam covers the documented boundary case.
• Dask chunk boundaries handled correctly: 2x1 and 2x2 mosaics with sub-tile chunks both match eager.
• No premature materialization: compute() only fires after the multi-block dask graph is built.
• [n/a] Benchmark exists or is not needed: test-only PR.
• [n/a] README feature matrix updated: no public API change.
• Docstrings present and accurate.

[brendancol]

Follow-up review (post-fixes)

Disposition for the four findings from the previous review:

• Fixture tempfile.mkdtemp -> tmp_path subdirs. Fixed.
• _assert_attrs_ok now checks transform[2] / transform[5]. Every call site passes the fixture origin; the windowed call passes the shifted origin (ox + _PIXEL_W*c0, oy + _PIXEL_H*r0). Fixed.
• Unused w_total removed; the windowed test now binds only h. Fixed.
• Dask multi-band test added (test_dask_multiband_2x1_mosaic_matches_eager) so per-block band handling is covered. Fixed.

Total: 7 tests pass locally. No deferrals, no dismissals.