Fix/dtype narrowing

[fishidaho]

Size index arrays by the axis they address, not the array beside them

What's wrong

indices is the only nnz-sized array in the VCS layout, and nothing ever picked its dtype. Instead, every path just inherited whatever the input carried:

• _construct.compress (src/vsparse/_construct.py) copied minor_indices.dtype straight into its output buffer.
• write_ivcs_elem (src/vsparse/_io.py) recorded np.dtype(v.indices.dtype).name, so a wide in-memory array stayed wide on every future read.

scipy hands out int64 indices for any array with enough nonzeros, so a 33k-gene minor axis was routinely stored at 8 bytes per nonzero instead of 4. This was found to be the case in memory, on disk, and in every kernel that walks it.
_rapid_load._filter_and_compact had the same conflation in a sharper form. One idx_dtype, keyed off nnz, was applied to two arrays with completely different bounds:

idx_dtype=np.int64ifindices.shape[0] >np.iinfo(np.int32).maxelsenp.int32new_indptr=np.zeros(..., dtype=idx_dtype) # indexed by nnz: needs int64 at scaleout_indices=np.empty(nnz_filtered, dtype=idx_dtype) # gene indices: never does

What this changes

Adds _indexutils.smallest_index_dtype(n) as the single place the rule lives and applies it at each point an index array gets sized:

Site	Bound it's keyed off
_construct.compress (new n_minor arg)	minor-axis length
_VCSBase.__init__	minor-axis length
write_ivcs_elem	minor-axis length
_filter_and_compact → new_indptr	total nnz
_filter_and_compact → out_indices	kept-gene count
transpose_major	(already did this inline; now uses the helper)

Verification

tests/test_index_dtypes.py (new, 58 cases across the shared shape fixtures):

• the boundary rule itself, exactly at INT32_MAX and one past it;
• from_scipy on an int64-indexed input produces int32 indices, and indices.nbytes == 4 * nnz;
• a minor axis genuinely larger than INT32_MAX still gets int64 (tiny nnz against a huge shape, so the test stays cheap);
• out-of-bounds indices still raise instead of wrapping;
• vcsc/ivcsc write→read round trips preserve values and shape while coming back narrow;
• the packed writer re-derives the dtype for an array that was handed to it holding int64;
• _filter_and_compact picks int32 for both when both fit, and keeps gene indices at int32 when the nnz-keyed half is forced to int64 — the regression case, since allocating a >INT32_MAX-nonzero matrix isn't testable directly.

Full suite: 819 passed, 37 skipped. ruff check src tests clean; ty check
introduces no new diagnostics (the one pre-existing no-matching-overload
on the read path is untouched).

Before/after on the headline path:

OLD from_scipy indices dtype: int64 -> 8 bytes/nonzero
NEW from_scipy indices dtype: int32 -> 4 bytes/nonzero

[fishidaho]

All of the active PRs right now are essentially load-bearing for all of the issues I have open right now. I think merging these in prior to an initial v0.1.0 PyPi release would be ideal.

None of them actually change the existing API; however, the future issues will likely touch the API or introduce new API function.