Add select(recalculate=False): a lazy, stats-preserving row/col selection

[aarmey]

Summary

select() previously always recomputed statistics fresh for the selected
sub-array. A caller that instead wants to evaluate a fixed (e.g.
train-derived) normalization against a different slice had no lazy option
for that: bracket indexing (__getitem__) keeps the parent's statistics,
but is documented to always eagerly materialize the selection as a dense
ndarray (toarray()[key]) -- fine for a small window, but not for a
selection covering a large fraction of a huge array.

This surfaced while getting BAL-Pf2's BiCV rank-selection sweep working at
its real scale (1.3M cells): scrise.rank_selection._bicv_trial scores a
model fit by indexing a normalized view with a boolean cell mask covering
~50% of all cells (X_mat[test_cell_mask]), which -- via __getitem__ --
eagerly builds a dense array on the order of tens of GB, for something that
should stay a lazy sparse view all the way through to the eventual
dense-times-sparse product.

Changes

select(rows, cols, *, recalculate=True) gains the recalculate keyword.
recalculate=False reuses this view's existing a/b/c/s (row_scale
sliced by rows -- exact, since it's a per-cell quantity; the per-gene
statistics sliced by cols, unaffected by which rows are selected) via the
existing from_stats() classmethod, rather than recomputing from the
selected sub-array's own data. The default (recalculate=True) is
unchanged.

The result is a real view (not an array), so it composes with @/
toarray() and stays lazy regardless of selection size, unlike
__getitem__ -- the caller decides when (or whether) to ever materialize
it, and even then the O(nnz) machinery is bounded rather than allocating a
dense block sized to the selection up front.

Test plan

• uv run pytest -- 1246 passed, 49 skipped, 3 pre-existing failures
  unrelated to this change (identical on main without this diff)
• uv run ruff check . / uv run ruff format --check .
• uv run ty check src/
• New tests in tests/test_norm_selection.py: matches __getitem__'s
  values, returns a genuinely lazy view that still composes with @,
  composes with a column selection too, and the default
  (recalculate=True) is unaffected by the new keyword.