Add an array-API median fallback for subtract_overscan

[mwcraig]

What

subtract_overscan calls xp.median(overscan.data, axis=overscan_axis)
(ccdproc/core.py), but median is not part of the array API standard, so
it raises AttributeError on backends such as array-api-strict that don't
provide it (numpy, jax and dask all have a native median and are
unaffected).

This adds ccdproc._nanfuncs.median, a small wrapper around the existing
nanmedian fallback, and uses it from a new _median_fallback helper in
core.py that tries the namespace's own median first and only falls back
to the array-API-only implementation when it's missing -- mirroring the
existing _percentile_fallback pattern already used for percentile.

Why

Part of #971 (array-API migration tracker). median joins nansum /
nanmean / nanstd / nanmedian (#906, #986) as another reduction that
isn't part of the array API standard and needs a pure-array-API fallback.

NaN semantics

numpy.median propagates NaN: if any value in the reduced slice is NaN, the
result is NaN. nanmedian does the opposite -- it ignores NaNs. Since
subtract_overscan's median fallback needs to match numpy.median
(not nanmedian), the new median function wraps nanmedian and then
restores NaN-propagation with a final xp.where(xp.any(xp.isnan(x), ...)).
On NaN-free input median is bit-for-bit identical to nanmedian, since it
reuses the exact same sort-and-average algorithm.

Testing

• ccdproc/tests/test_nanfuncs.py: added median (against np.median) to
  the differential _FUNCS table, which already exercises NaN-scattered,
  all-NaN, single-non-NaN, integer and boolean inputs across a range of
  shapes and axes -- this is what pins the NaN-propagating semantics. Also
  added median to test_no_warning_on_all_nan_slice and test_bad_axis
  (verified numpy.median, unlike numpy.nanmedian, does not warn on
  all-NaN input).
• array-api-strict: all "module 'array_api_strict' has no attribute
  'median'" errors are eliminated -- confirmed zero occurrences anywhere in
  the suite after this change.
• numpy backend: ccdproc suite is unaffected (0 failures, native xp.median
  path taken, byte-identical to before).
• dask backend: test_nanfuncs.py + test_ccdproc.py pass.

Note on the strict-suite failure count: the 4 tests originally cited as
failing due to this bug (test_ccd_process[*],
test_ccd_process_gain_corrected) do not go green with this fix alone.
Diffing the full array-api-strict failure list before/after shows the same
85 tests fail both times -- but for 3 of these 4 tests the reason changes,
from AttributeError: ... has no attribute 'median' to two separate,
pre-existing bugs that this fix now uncovers (they were previously masked
because subtract_overscan raised first):

• ccd_process's bad-pixel-mask handling uses xp.asarray(bad_pixel_mask, dtype=bool) with the Python builtin bool instead of the namespace's
  xp.bool (core.py, in ccd_process), which array-api-strict rejects.
• gain_correct builds its gain array with xp.asarray(gain_value) without
  specifying device=, so it ends up on the default device while the image
  data is on array-api-strict's non-default device, and the subsequent
  multiply fails with "Arrays from two different devices".

Both are real, independent array-API incompatibilities, out of scope for
this PR, and worth filing as separate follow-up issues against #971.

🤖 Generated with Claude Code

https://claude.ai/code/session_01RQMJZUaaxfqGDk41GLSaFK

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 97.44%. Comparing base (3e644d8) to head (5399fc2).
⚠️ Report is 9 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #989      +/-   ##
==========================================
+ Coverage   97.42%   97.44%   +0.01%     
==========================================
  Files           9        9              
  Lines        1749     1760      +11     
==========================================
+ Hits         1704     1715      +11     
  Misses         45       45              

Flag	Coverage Δ
dask	96.57% <100.00%> (+0.02%)	⬆️
jax	96.63% <100.00%> (+0.02%)	⬆️
numpy	97.32% <100.00%> (+0.01%)	⬆️

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.