Keep BiCV's held-out cell blocks lazy for duck-typed backends

[aarmey]

Summary

_bicv_trial evaluates a rank's fit by row-restricting X.X with a boolean
cell mask (X_mat[train_cell_mask]/X_mat[test_cell_mask]) and handing the
result to rmatmul/calc_W, deliberately to avoid materializing the raw
data (see the existing comment: "reaches the raw data through products...
rather than materialising a block of it"). That holds for a plain ndarray
or scipy-sparse X_mat, but not for a vsparse normalized view (e.g.
BAL-Pf2's lazy-normalized-view AnnData): bracket indexing such a view is
documented to always eagerly build a dense ndarray for the selection.

At BAL-Pf2's real scale (1.3M cells), a single train/test split (~50% of all
cells) would materialize on the order of tens of GB, once or twice per BiCV
trial, across every rank/repeat in a 100-rank sweep. This never surfaced
before because an unrelated vsparse memory issue (fixed in
meyer-lab/vsparse#49) always killed these runs earlier in the pipeline,
during compression -- this only became visible once that was fixed and a
real subsample run got far enough to reach the evaluation step.

Changes

• _restrict_rows(X_mat, mask): uses vsparse's new
  select(recalculate=False) (Add select(recalculate=False): a lazy, stats-preserving row/col selection vsparse#50) to stay a genuinely
  lazy view for a duck-typed backend, falling back to ordinary bracket
  indexing for plain dense/sparse X_mat (unaffected, still cheap for
  those). Used at all three X_mat[mask] sites in _bicv_trial.
• _test_block_moments gains a third branch for the same duck-typed case:
  selects the cell subset lazily, then streams it in bounded row chunks
  (_MOMENT_CHUNK_BUDGET_BYTES) rather than ever materializing the whole
  subset as one dense block -- mirroring the existing scipy-sparse branch's
  own streaming design, just against a different backend's primitives.
• pyproject.toml: temporarily pins vsparse to its (as yet
  unmerged/unreleased) select-without-recalculate branch -- drop once
  Add select(recalculate=False): a lazy, stats-preserving row/col selection vsparse#50 merges and releases, reverting to the plain PyPI
  version constraint.

Test plan

• uv run pytest scrise/tests/ -- 131 passed, 1 skipped (pre-existing
  skip, unrelated)
• uv run ruff check . / uv run ruff format --check .
• uv run ty check scrise/
• New tests in test_rank_selection.py: _restrict_rows matches
  bracket indexing for a normalized view while staying lazy (both VCSR
  and VCSC), is a no-op passthrough for plain dense/sparse arrays,
  _test_block_moments's new branch matches a dense reference, streams
  across multiple row chunks correctly, and the streaming loop's own
  peak memory is bounded by the chunk size rather than the selected
  block's total size (isolated from the one-time, pre-existing cost of
  building the lazy selection itself, which is a separate, bounded,
  single-occurrence cost this PR doesn't change).

Related

• Replace chunked-transpose matmul fallback with a native minor-axis kernel vsparse#49 (chunked-transpose matmul fallback -> native
  minor-axis kernel)
• Add select(recalculate=False): a lazy, stats-preserving row/col selection vsparse#50 (adds select(recalculate=False), which this PR
  depends on)