Add a per-row/per-column statistics sidecar to VCS arrays

[aarmey]

Implements the statistics sidecar proposed in #26, shaped around the normalization recipe from #40 (y = (g(x·a·b) − c)·s).

What's in the sidecar

VCSCArray/VCSRArray now carry arr.stats, a vsparse.VCSStats. It is computed on first access and then kept:

field	axis	meaning
row_sums	cells	total counts (obs["total_counts"])
row_nnz	cells	stored nonzeros (n_genes_by_counts; explicit zeros not counted)
col_sums	genes	total counts
col_nnz	genes	n_cells_by_counts
col_depth_sums	genes	Σᵢ x[i,j] / row_sums[i]

These are recipe-independent, and they determine a and b for every recipe:

• a = target / row_sums, where target is a constant or median(row_sums).
• b = 1 / (target · col_depth_sums), or 1 / col_sums when there's no depth normalization, because target factors out of the sum.

Normalization now always takes a and b from the sidecar, so a view built from a restored sidecar matches a freshly computed one exactly. That makes raw and cp10k_log1p views free once the sidecar exists. c and s are moments of the nonlinear g, so there is no recipe-independent summary of them, and they still take one pass per recipe (for VCSC that pass no longer computes b too).

The stats live on the plain array, so they exist (and get written) whether or not a normalized view was ever built.

On disk

• write_h5ad/write_zarr always write the sidecar as X/stats (and raw_X/stats), plus arrays nested in uns through the registered VCSC/VCSR codec. The group carries encoding-type: vsparse-stats and encoding-version: 0.1.0.
• read_h5ad/read_zarr/load_packed attach it when it's usable. On read it degrades gracefully:
  • A missing sidecar (older files, other tools), an unknown version, or a missing field is silently ignored.
  • A sidecar whose shape doesn't match the array is discarded with a UserWarning.
  • In both cases the array recomputes the stats only when they're first needed.
• Older vsparse versions read the arrays explicitly by key, so they ignore the extra subgroup.

Numerical notes

• b is now computed as target · Σ x/depth rather than Σ x/(depth/target). Results differ at the ulp level; all existing tolerance-based tests pass.
• IVCS packing sorts each value group's indices. To keep a sidecar written before a round trip bit-identical to one recomputed after it, VCSC depth sums add each group in row order. _compress now uses a stable sort, so groups built from scipy input are usually already sorted and the sort is skipped.

Not in this PR

• Wiring Design a composable lazy Selection object for VCSCAnnData #25's Selection to filter from the sidecar (that piece of Create a statistics sidecar as an on-disk format feature #26's acceptance criteria). arr.stats exposes everything needed for it.
• Carrying stats through subsetting or .T. Subsets recompute lazily; full-slice copies share the parent's stats.

Checks

• pytest: 1615 passed (27 new in tests/test_stats_sidecar.py). They cover agreement with dense recomputation (Hypothesis), exact round trips for h5ad/zarr × VCSC/VCSR (X, raw_X, nested in uns), load_packed, invariance to order within groups, every degrade-gracefully path, and a check that restored stats are never recomputed.
• ruff, codespell, ty: clean. Fast benchmarks: no regressions.
• xenon/radon: no new C-or-worse blocks; _column_stats_major_is_col dropped out of rank C. jscpd: no new clones. vulture: only a false positive on numpy's flags.writeable.

https://claude.ai/code/session_01QBeTZy6BKKGKGJ7fwvkWgM