Reject zero-denominator RATIONAL/SRATIONAL tags (#2313)

[brendancol]

Closes #2313.

Summary

• _read_value in xrspatial/geotiff/_header.py used to coerce a malformed RATIONAL or SRATIONAL value (denominator=0) to 0.0. It now raises ValueError with the offending tag and denominator in the message, matching the actionable failure style from PR Pin actionable failure modes for unsupported COG writer inputs (#2286 prod-ready wave B) #2301.
• The tag id is threaded through the three call sites in parse_ifd, so the message can name XResolution / YResolution / ResolutionUnit directly when it's one of those.
• The two existing tests that pinned the old silent behaviour are flipped to assert the new failure, and a new regression test (test_rational_zero_denominator_2313.py) builds a TIFF whose XResolution rational has a zero denominator and confirms both the header parser and open_geotiff fail loudly.

Backend coverage

Header decoding is shared across all four backends (numpy, cupy, dask+numpy, dask+cupy), so the change applies uniformly. No backend-specific code is touched.

Test plan

• pytest xrspatial/geotiff/tests/test_header.py
• pytest xrspatial/geotiff/tests/test_rational_zero_denominator_2313.py
• pytest xrspatial/geotiff/tests/ -k "not gpu and not cuda" (4282 passed, 51 skipped)

[brendancol]

PR Review: Reject zero-denominator RATIONAL/SRATIONAL tags (#2313)

Blockers

None.

Suggestions

• test_header.py:261 and test_header.py:272: the two new tests do from xrspatial.geotiff._header import TAG_X_RESOLUTION (and Y) inside the method body. The module is already imported at the top of the file (_read_value, parse_all_ifds, parse_header); folding the tag constants into that same import keeps the block consistent and avoids the local import.

Nits

• _header.py:88-96: _TAG_NAMES covers only X/YResolution and ResolutionUnit, which is the right scope for this PR. If a follow-up threads tag= through other error sites, growing this map will help. Flagging it so it doesn't get forgotten.
• test_rational_zero_denominator_2313.py:147-208: test_yresolution_zero_denominator_named_in_error reimplements most of _build_tiff_with_rational_xres. Adding a which: int = 282 parameter to the helper (default XResolution, 283 for YResolution) would let the YResolution test reuse the helper instead of duplicating ~60 lines.

What looks good

• Error message follows the "refusing to parse possibly malformed TIFF" pattern already used by parse_ifd (lines 666/672 in the original).
• All three _read_value call sites in parse_ifd get tag=tag, including the pre-scan path that swallows ValueError. That swallow is safe because _DIMENSION_TAGS doesn't include any RATIONAL tag.
• The regression test exercises both parse_all_ifds and open_geotiff, so the failure is verified at the public entry point, not just the internal helper.
• The two existing tests that pinned the old silent behaviour are flipped to assert the new failure, not deleted. A regression that reintroduces the silent path would fail loudly.
• BigTIFF coverage is implicit: parse_ifd uses the same _read_value for both classic and BigTIFF entries.
• Error message includes the element index, so a malformed entry inside a multi-element RATIONAL array gets pinpointed.

Checklist

• Algorithm matches spec (TIFF treats denom=0 as malformed)
• NaN handling unchanged (rationals never produced NaN before either)
• Edge cases covered (count>1 in test_rational_denominator_zero_in_array_raises)
• No premature materialization or unnecessary copies
• Benchmark coverage: not needed (error path)
• README feature matrix: not applicable (pure bug fix)
• Docstrings present and accurate (_read_value docstring documents the new behaviour)

[brendancol]

Follow-up review after ca86979

Re-checked the two items that needed action and confirmed the new commit addresses them.

Disposition

• Suggestion (hoist local imports in test_header.py): fixed in ca86979. TAG_X_RESOLUTION and TAG_Y_RESOLUTION are now in the top-level import block, and the two local imports inside the test methods are gone.
• Nit (share helper between XResolution and YResolution tests): fixed in ca86979. _build_tiff_with_malformed_resolution now takes a which parameter, and test_yresolution_zero_denominator_named_in_error is a one-liner that reuses the helper. Net -41 lines in the test file.
• Nit (_TAG_NAMES map could grow): deferred. The map only exists to label tags in error messages produced by _read_value, and the only error sites that take a tag= kwarg today are the RATIONAL / SRATIONAL branches. Adding entries for tags that no error site references would be dead code. If a future PR threads tag= through other error sites, growing the map should be part of that PR.

Tests: pytest xrspatial/geotiff/tests/test_header.py xrspatial/geotiff/tests/test_rational_zero_denominator_2313.py -> 44 passed.

No new findings on this pass.