Forward allow_rotated/allow_invalid_nodata to VRT per-source reads (#2672)

[brendancol]

Closes #2672

The public read_vrt accepts allow_rotated, allow_unparseable_crs, and allow_invalid_nodata, and documents the first and third as opt-ins forwarded to the per-source GeoTIFF reads. They weren't actually forwarded. The eager internal call and the chunked per-chunk reader only passed the codec flags down, so a caller who set allow_rotated=True or allow_invalid_nodata=True still hit a typed rejection on a rotated source TIFF or a source with non-representable integer nodata. Under missing_sources='warn' that rejection got reclassified as a hole, silently dropping the tile.

What changed:

• Added allow_rotated and allow_invalid_nodata to the internal read_vrt in _vrt.py and forwarded them to the per-source read_to_array call.
• Forwarded both opt-ins from the eager backend call and from _vrt_chunk_read (the chunked per-chunk reader) in _backends/vrt.py.
• allow_unparseable_crs is a validation-layer opt-in only. read_to_array doesn't accept it, and the VRT takes its CRS from the already-validated VRT XML, so it doesn't need per-source threading.

Backend coverage: numpy (eager) and dask+numpy (chunked) VRT read paths. The change lives in the shared per-source read plumbing, so the cupy / dask+cupy VRT paths inherit it through the same internal read_vrt.

Test plan:

• New tests in xrspatial/geotiff/tests/vrt/test_source_opt_ins_2672.py build a single-source VRT over a rotated source TIFF and over a uint16 source with NaN GDAL_NODATA. Each case is rejected without the opt-in and reads cleanly with it, on both eager and chunked paths.
• A missing_sources='warn' test confirms the opt-in keeps a readable source out of the hole bucket instead of emitting a fallback warning and recording a false hole.
• Reverting the source-side forwarding makes the five "with opt-in" cases fail, which confirms the tests exercise the fix.
• Full VRT suite (491 tests) plus the read nodata/georef suites pass; flake8 clean on the changed files.

No CHANGELOG edit (updated separately at release time). No new public API: the opt-ins already existed and were documented; this makes them take effect.

[brendancol]

PR Review: Forward allow_rotated/allow_invalid_nodata to VRT per-source reads (#2672)

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None. The two opt-ins now reach read_to_array on both the eager (_vrt.py per-source read) and chunked (_vrt_chunk_read) paths, and the GPU / dask+GPU VRT paths inherit it because the per-source decode is always the same CPU read_to_array.

Nits (optional improvements)

• _vrt_chunk_read in _backends/vrt.py gained allow_rotated / allow_invalid_nodata params with no docstring lines, but the existing codec kwargs there are also undocumented, so it matches the surrounding style. A one-line note would mirror the public docstring; optional since the function is internal-only.

What looks good

• allow_unparseable_crs is correctly left out of the per-source read: read_to_array doesn't accept it, and the VRT CRS comes from the already-validated VRT XML. This matches the dask backend, which also keeps allow_unparseable_crs at the validation layer and forwards only allow_rotated / allow_invalid_nodata to the reader.
• Both internal read_vrt call sites (eager at _backends/vrt.py:540, chunked at :788) are updated; nothing was missed.
• The tests prove the fix, not just exercise it. Reverting the source-side forwarding flips the five "with opt-in" cases to failures, and they pass with it. Both rejection types (RotatedTransformError, InvalidIntegerNodataError) are real ValueError subclasses, so the missing_sources='warn' hole-reclassification test hits the actual catch path described in the issue.
• Temp filenames carry the issue number (rotated_src_2672.tif, etc.), and the tests reuse the existing _write_rotated_tiff / _build_uint16_tiff_1774 builders so the byte layout stays shared with the non-VRT reader tests.

Backend coverage note

The per-source read is CPU-only (read_to_array) for all four VRT output backends, so numpy / dask+numpy are directly tested and cupy / dask+cupy inherit the same fix. A dedicated cupy test would need a GPU runner; the eager/chunked coverage is the right level here.

Checklist

• Algorithm matches reference/paper: n/a (plumbing fix)
• All implemented backends consistent: yes (shared per-source read)
• NaN handling correct: unchanged; nodata masking path untouched
• Edge cases covered: yes (rotated source, NaN integer nodata, warn-mode hole)
• Dask chunk boundaries: yes (opt-ins forwarded per chunk task)
• No premature materialization or copies: yes
• Benchmark: not needed (no perf-sensitive change)
• README feature matrix: not needed (no new function, no backend change)
• Docstrings: public docstring already documented the forwarding; the fix makes it true

[brendancol]

Follow-up review (after nit fix)

The one nit from the previous pass is addressed: _vrt_chunk_read now documents that allow_rotated / allow_invalid_nodata are forwarded to the internal reader so each task's per-source GeoTIFF read honors the opt-ins, matching the eager path.

No new findings. No Blockers, no Suggestions. The 9 tests in test_source_opt_ins_2672.py still pass and flake8 is clean on the changed file. Nothing else to fix.