Fix array-namespace dtype handling in ccd_process, transform_image, combine, and ImageFileCollection

[mwcraig]

Part of #971 (the "namespace dtypes" bucket).

Strict-backend count on top of main 711bb26: 41 failed -> 35 failed (456 -> 463 passed; 43 xfailed, 5 xpassed unchanged). No test regresses on strict, numpy, jax, or dask.

ccd_process, transform_image, combine(), and ImageFileCollection.ccds()/data() all leaked NumPy-specific dtype conventions across the array-namespace boundary: a builtin bool instead of xp.bool for ccd_process's bad-pixel mask, running a transform function directly on a boolean mask in transform_image, and handing a NumPy dtype object (rather than a plain array) to xp.asarray(..., dtype=...) when converting FITS data to a requested namespace in combine() and ImageFileCollection. Each is fixed to speak the array API idiom already used elsewhere in the codebase (xp.bool, casting the mask to the data dtype before transforming, converting to native byte order and letting xp.asarray do a plain conversion). A new _native_numpy helper in core.py, next to _is_array, is shared by combine() and ImageFileCollection. ImageFileCollection's mask conversion also had to switch from the public .mask setter to the private _mask attribute, since CCDData's public setter unconditionally converts its value back to NumPy - the same workaround ccd_process already used, with a TODO to remove it once CCDData supports array namespaces.

Fixing ccd_process's mask dtype uncovered a second, previously-masked bug: log_meta.py's _replace_array_with_placeholder (used by the @log_to_metadata decorator to build the auto-logging string) crashed for a bare, non-NDData array-API array that does not implement __len__ - which array_api_strict deliberately does not, per the array API standard. It fell through to value.data, which such an array does not have, raising AttributeError instead of the TypeError the code was written to catch. That crash was masking ccd_process's tests even after the mask-dtype fix, so it is fixed here too (broadened the except to also catch AttributeError).

Test-body fixes bundled with the corresponding source fix: test_transform_image discarded the return value of xpx.at(...).set(...) (a silent no-op on JAX/strict, the same class of bug as #963) and did not build a real boolean mask; both are fixed, and the test now asserts the transformed mask is True only at the expected pixel. test_ccd_process/test_ccd_process_gain_corrected built their bad-pixel mask as a float array and compared it against the (now boolean) result with xpx.isclose, which strict namespaces reject for non-numeric dtypes; both now build a boolean mask and compare with ==. A new test_generator_ccds_converts_to_array_namespace in test_image_collection.py exercises the same FITS-to-namespace conversion in ImageFileCollection.ccds() that combine() already had a regression test for (A4 in the working plan; not on the original strict-failure list since no existing test passed array_package= to ImageFileCollection, but the code path is byte-for-byte the same pattern as combine(), so it is included here).

One target test remains failing on strict for a reason outside this PR's scope: test_combine_ccd_with_uncertainty_and_mask_from_fits[function] reaches a second, unrelated bug once this PR's fix to combine() clears the first one - test_combiner.py's _make_mean_scaler calls ccd_data.data.mean() as a method, which array_api_strict arrays do not support (only the top-level xp.mean function form is part of the standard). That is a bug in test_combiner.py, a file intentionally left untouched here to avoid conflicting with a parallel PR (fixing Combiner.clip_extrema) also touching that file; [mean] (the other parametrization) passes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01S36ZzAAVXVm32vuTdtCQME

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 97.75%. Comparing base (8f36710) to head (f85764e).
⚠️ Report is 9 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #995      +/-   ##
==========================================
+ Coverage   97.47%   97.75%   +0.28%     
==========================================
  Files           9        9              
  Lines        1781     1785       +4     
==========================================
+ Hits         1736     1745       +9     
+ Misses         45       40       -5     

Flag	Coverage Δ
dask	96.90% <100.00%> (+0.28%)	⬆️
jax	96.96% <100.00%> (+0.28%)	⬆️
numpy	97.64% <100.00%> (+0.28%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.