BUG: ma.__array_wrap__ no longer propagates a dtype-invalid fill_value - #32438

Description

When a ufunc changes the output dtype of a MaskedArray (e.g. np.strings.find applied to a string masked array returns int64 match positions), the result inherits the raw _fill_value from its input via _update_from inside MaskedArray.__array_wrap__, without any validation against the new dtype.

That inconsistent state does not surface immediately — repr, indexing, etc. all work — but the next .view() re-runs _check_fill_value through __array_finalize__ and raises:

TypeError: Cannot convert fill_value N/A to dtype int64

Minimal reproducer (from gh-32401):

import numpy as np
col = np.ma.MaskedArray(['foo', 'bar', 'baz'], mask=False, fill_value='N/A')
result = np.strings.find(col, 'foo')   # dtype changes str -> int64
result.view(np.ma.MaskedArray)         # TypeError before this patch

This also breaks astropy's MaskedColumn, whose .data property calls .view(np.ma.MaskedArray) internally on every __getitem__ (astropy/astropy#20257).

Fix: after _update_from in __array_wrap__, if the result dtype differs from the input dtype, revalidate the copied _fill_value with _check_fill_value:

• a fill value that is convertible to the new dtype is kept (converted), so e.g. a numeric fill value survives an str -> int transition;
• a fill value that cannot be represented is dropped, so the output falls back to its default fill value instead of carrying an invalid one;
• same-dtype ufuncs are untouched (the check is gated on result.dtype != self.dtype), preserving all existing behavior.

Testing

Two regression tests added to numpy/ma/tests/test_core.py::TestUfuncs:

1. test_dtype_changing_ufunc_does_not_inherit_invalid_fill_value — the issue's reproducer; asserts the invalid fill value is not carried over and .view(MaskedArray) no longer raises. Fails on unpatched main with the reported TypeError.
2. test_dtype_changing_ufunc_keeps_convertible_fill_value — asserts a numeric fill value on a string array survives a str -> int64 ufunc (converted, not discarded).

Full suite with the patch applied:

numpy/ma/tests/test_core.py      4172 passed, 1 skipped, 2 xfailed
numpy/ma/tests/test_extras.py    }
numpy/ma/tests/test_regression.py } 185 passed
numpy/ma/tests/test_mrecords.py  }

Notes for reviewers

• The failure mode is "stale state discovered late", so I chose revalidation at wrap time over raising eagerly at wrap time — an eager raise would turn previously-working code paths (any dtype-changing ufunc on an array that happens to carry a fill value) into hard errors. Dropping to the default fill value keeps those paths functional while guaranteeing internal consistency.
• An alternative considered was validating inside _update_from itself, but that function is used by many non-ufunc callers (__array_finalize__, explicit attribute copies) where silently dropping the caller-provided fill value would be more surprising.