Check VCS index bounds in one parallel pass

[aarmey]

Summary

_VCSBase.__init__ checks that every index is in [0, n_minor) with indices.min() and indices.max(): two single-threaded passes over the whole array. On large files that is a large share of the load time, spent after the indices were already decoded in parallel.

• New _ops.indices_in_bounds() checks arrays of at least 2²⁰ signed-integer indices with one parallel numba pass. Smaller or unsigned arrays keep the numpy min/max path.
• The constructor uses it. Behavior is unchanged: an out-of-range index still raises ValueError before the dtype narrowing, so it can't wrap silently.

Results

full_IBD_cohort (1.19M cells, 3.6 billion indices = 14.6 GB of int32), warm cache:

	Before	After
Bounds check in the constructor	~1.2 s (min 0.6 s + max 0.6 s)	one parallel pass
load_packed	5.30 s	3.50 s

What's left is mostly decoding the packed indices (~2.9 s).

Notes

• This doesn't change index dtypes. Decoding straight into uint16 would halve the memory, but it would break the int32-minimum index policy from Fix/dtype narrowing #29 that the kernels rely on, so I left it out.

Test plan

• test_parallel_bounds_check: the parallel path agrees with min/max for int32/int64 arrays with no bad index, a negative index, one too large, and 2**31 - 1.
• test_large_out_of_bounds_index_raises: the constructor raises when the parallel check finds a bad index (VCSC and VCSR).
• Full suite: 1588 passed, 71 skipped. ruff check, ruff format --check, codespell and ty check are clean; vulture, radon and xenon report nothing new.

https://claude.ai/code/session_01QBeTZy6BKKGKGJ7fwvkWgM