Design a composable lazy Selection object for VCSCAnnData

[fishidaho]

VCSCAnnData.__getitem__ (src/vsparse/_anndata_class.py:154-182) is always an eager copy. This means every subset operation pays a full materialization cost, and there's no way to compose several filters (e.g. an obs mask, then a var mask, then a further obs mask) without paying for each step separately.

This matters for the target end-to-end experience as a chain like:

va = va[va.obs["condition"] == "control", :]
va = va.sample_obs(n=5000, by="patient", random_state=42)  # ISSUE-28
X = va.normalize()                                          # ISSUE-29

should compose cheaply and only touch data once, at the point something
actually needs values.

Proposed scope

• A Selection object that records row/column filters compositionally without touching data until something actually needs values.
• VCSCAnnData gains a lazy indexing path that returns a Selection-backed view instead of an eager copy; .compute()/.materialize() (naming TBD) forces it when a caller actually needs a real VCSCAnnData.
• Once IVCSC/IVCSR on-disk format with a byte-offset index (proposed v2 format) #24 's byte index exists, filters expressible entirely in terms of obs/var metadata should resolve to zero reads until materialization.
• Decide how this interacts with the NormalizedView fix as a Selection composed before constructing a NormalizedView is the clean way to guarantee post-selection statistics.

Acceptance criteria

• A multi-step subsetting chain (e.g. six filters in sequence) composes in under a millisecond with no data read.
• Materializing a composed Selection performs exactly one read, not one per step.
• Existing VCSCAnnData indexing behavior (obs/var labels, ints, slices, boolean masks) is preserved for callers that don't opt into laziness.

References

• src/vsparse/_anndata_class.py:154-182 (class docstring, __getitem__).