Fix bytes codec endian for big-endian structured dtypes

[emfdavid]

What

convert_to_codec_pipeline chose the array→bytes codec's byte order from dtype.byteorder.
numpy reports "|" for every structured dtype regardless of what its fields hold, so
big-endian record data fell through to the little-endian BytesCodec() default. zarr then
faithfully decodes the referenced bytes in the byte order it was handed — the wrong one — so a
virtual reference to a FITS BinTableHDU (e.g. an SDSS spectrum) reads back silently wrong
values. No error, just garbage.

Simple scalar arrays were unaffected (np.dtype('>f4').byteorder == '>'), which is why this
only ever surfaced on structured/record dtypes.

Fix

Derive the byte order from the fields rather than from the dtype itself:

• walk dtype.fields, looking through subarray fields (whose order hides behind .base) and
  nested structs (hence the recursion);
• ignore fields that carry no byte order at all — single-byte numbers, strings — instead of
  letting them imply little-endian;
• a structured dtype whose fields mix byte orders now raises ValueError: a Zarr V3 array
  has a single bytes codec and so cannot express per-field byte order. FITS is uniformly
  big-endian by standard, so this doesn't bite the motivating case; happy to downgrade to a
  warning if you'd prefer.

numpy gotchas the naive dtype.byteorder == ">" check misses, all covered by the new helper and
tests:

dtype	.byteorder
np.dtype([('flux','>f4')])	`'
np.dtype('<f4')	'=' — not '<'
('>f4', (3,)) subarray field	`'
nested struct	`'
S8, i1	`'

Companion zarr fix

This PR is one half of the story. It makes VirtualiZarr write endian: big for a big-endian
struct; zarr then has to honor it when decoding the referenced bytes. That decode-side branch
for structured dtypes landed in zarr-developers/zarr-python#4142 (merged; ships in zarr
3.2.2). Before that fix, zarr ignored the codec's endian for structs, so even a
correctly-tagged array read back wrong. Both halves are needed for a virtual big-endian struct to
read correctly end-to-end.

Testing

• new test_codecs.py cases covering the dtypes above (incl. the mixed-order raise); the codec
  pipeline now emits BytesCodec(endian="big") for a big-endian struct. These pass on released
  zarr 3.2.1 (what CI pins);
• end-to-end (FITS BinTableHDU → kerchunk → VirtualiZarr → Icechunk → native read
  byte-identical to astropy, no .view('>f4') consumer workaround) additionally requires zarr's
  decode fix, i.e. a zarr carrying fix: byte-order handling for structured dtypes in the bytes codec zarr-python#4142 (main today, 3.2.2 on
  release).

Sibling to #1037 (the list/base64 struct-ref fix) — same FITS-binary-table motivation.

Heads-up: a separate, pre-existing break (not this PR)

On zarr main, virtualizarr/utils.py:130 does endian.value, but zarr changed
BytesCodec.endian from an Endian enum to a plain str, giving 4 test_kerchunk.py
failures. CI is green only because it pins released zarr; this will surface when zarr 3.2.2
releases. Flagging separately — glad to file an issue or fold a fix in if that's useful.

Acceptance criteria

• Tests added — new cases in virtualizarr/tests/test_codecs.py
• Tests passing — codec suite + FITS BinTableHDU end-to-end green locally on released zarr 3.2.1; CI here is authoritative
• No test coverage regression — this only adds tests
• Full type hint coverage — the new _byte_orders / _is_big_endian helpers are annotated
• Changes are documented in docs/about/releases.md
• New functions/methods listed under docs/api — n/a, _byte_orders / _is_big_endian are private internal helpers
• New functionality has documentation — n/a, this is a bugfix, not new functionality

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 90.41%. Comparing base (fbe8afc) to head (be8f65d).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1047      +/-   ##
==========================================
+ Coverage   90.36%   90.41%   +0.05%     
==========================================
  Files          37       37              
  Lines        2242     2254      +12     
==========================================
+ Hits         2026     2038      +12     
  Misses        216      216              

Files with missing lines	Coverage Δ
virtualizarr/codecs.py	94.36% <100.00%> (+1.14%)	⬆️

🚀 New features to boost your workflow:

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