Fix VCSCAnnData.copy() silently dropping X/raw_X

[aarmey]

Summary

VCSCAnnData stores X/raw_X in private _vcs_X/_vcs_raw_X attributes,
since anndata's own X validation rejects a VCSCArray/VCSRArray directly
(see the class docstring). It never overrides .copy(), though, so the
inherited anndata.AnnData.copy() copies the standard (unused, always-None
for this class) _X attribute instead:

>>> ad = VCSCAnnData(X=arr)
>>> ad.X is None
False
>>> c = ad.copy()
>>> type(c)
<class 'anndata.AnnData'>
>>> c.X is None
True

X silently comes back None, and the returned object is downgraded to a
plain AnnData rather than preserving VCSCAnnData (or a subclass, such as
one overriding the X property for a lazy-normalized view -- as BAL-Pf2's
own import_data() does). Any caller relying on adata[mask].copy() hits
this immediately -- in particular parafac2's BiCV rank-selection routine
does exactly this for its train/test splits, which is how this surfaced.

Adds an explicit copy() override that copies every field (mirroring
__getitem__'s existing field-by-field approach), including a real
VCSCArray/VCSRArray copy of X/raw_X, and returns type(self).

Test plan

• uv run pytest (1266 passed, 53 skipped, 1 pre-existing unrelated
  failure -- test_property_normalization.py::test_recipe_matches_reference[scanpy-VCSRArray],
  a numeric edge case on an all-constant column that reproduces
  identically on main without this change) -- includes new tests in
  tests/test_anndata_class.py covering: X/raw_X preserved and
  independent after .copy(), obs/var/uns preserved, a subclass
  overriding the X property round-tripping through .copy(), and the
  exact adata[mask][:, mask].copy() pattern parafac2's BiCV uses.
• uv run ruff check . / uv run ruff format --check . / uv run codespell
• uv run ty check