Swaps the order to the correct thing and thus

closes gh-23651

BUG: Fix masked array raveling when `order="A"` or `order="K"`

Giving a warning for invalid/overflow in settitem/casts is right
(IMO), however for masked arrays it can be surprising since the
warning is not useful if the value is invalid but also masked.

So, simply blanket ignore the relevant warnings in setitem via errstate.
(There may be some other cases like `.astype()` where it might be
helpful to MA users to just blanket opt-out of these new warnings.)

Closes gh-23000

This transitively fixes gh-22912.  I had alooked a bit into whether
it is worthwhile to preserve the mask order, but TBH, we seem to not
do so in so many places, that I don't think it really is worthwhile.

Applying `order="K"` or `order="A"` to the data and mask separately
is an big bug though.

Closes gh-22912

Fixes #22041

* add regression test and fix for creating a masked array with a
structured dtype; the test is simply for lack of error in the repoducer

* the concern expressed by core team in matching issue was that
`astropy` might be negatively affected; I ran full `astropy` (hash: `c9ad7c56`)
test suite locally with this feature branch and it seemed "ok,"
just 1 unrelated network failure in the network-requiring tests
(`test_ftp_tls_auto`):

```1 failed, 21430 passed, 3490 skipped, 176 xfailed, 23275 warnings in
430.18s (0:07:10)```

[skip cirrus]