Create a statistics sidecar as an on-disk format feature

[fishidaho]

This is somewhat load-bearing for #25's zero-read filtering gate

_rapid_load._cell_totals already proves the core trick works: a cell's total count is computable from major_ptr/values/value_ptr alone —
far fewer entries than nnz — without ever touching the packed indices stream. This generalizes: obs/total_counts and n_genes_by_counts-style per-cell/per-gene statistics can be derived the same way, and a two-predicate QC filter built on them touches X not at all.

Today that only works incidentally, when a particular file happens to carry scanpy-computed QC columns in obs/var — not because vsparse guarantees or maintains them. There's no first-class mechanism for #25 's selection algebra to rely on a sidecar being present, current, or complete.

Proposed scope

• Define a sidecar schema (per-cell and per-gene statistics — at minimum total counts, n-genes/n-cells-by-counts, matching what QC filtering typically needs) as a designed, versioned part of the write path, not something derived ad hoc.
• Writers populate it automatically on write_h5ad/write_zarr; readers can trust it's present and correct for files written by this version, and degrade gracefully (recompute) for older files that don't have it.
• Wire Design a composable lazy Selection object for VCSCAnnData #25's Selection object to serve obs/var-driven filters from the sidecar when available, entirely without a data read.

Acceptance criteria

• Every file written by write_h5ad/write_zarr carries the sidecar by default.
• Sidecar values match recomputation exactly.
• A Selection filter expressible purely in sidecar terms performs zero reads against X.

References

• src/vsparse/_rapid_load.py (_cell_totals).