Fix OOM in scm_gene_stat on large datasets (#98)

[aviezerl]

Fixes #98.

Chunking

quant was ceiling(nrow(mat_fg)/mc_cores), so the gene chunk size grew as mc_cores shrank. Each chunk densified three matrices (mat_fg, mat_ds, mat_n) at quant x n_cells x 8B each, multiplied by every fork.

Now chunks on a fixed ~50MB dense block, so peak memory per fork is independent of mc_cores. Fixed-size chunks can leave a 1-gene tail, so the three row subsets use drop=FALSE.

I did not cap the workers at 3 as the issue suggests - mc_cores is the user's setting, and with 50MB blocks the cap isn't needed.

sz_cor

apply(mat_fg[,subs], 1, cor) coerces the whole sparse submatrix to dense. Replaced with .sparse_row_cor, a sums-of-products Pearson using one sparse matvec, no densification. Chose this over chunking the apply because it also removes a per-gene sort over all cells.

Verification

Golden-master: gene_stat is identical() to the previous implementation on

case	shape	exercises
small	800 x 20,000	chunk count changes 4 -> 3
wide	300 x 50,001	the >50000 sz_cor branch
tail	313 x 20,000	1-gene tail chunk

.sparse_row_cor also matches apply(mat, 1, cor) to 3 decimals including the zero-variance -> NA case. Not covered: the OOM itself at 80k cells, which needs the real dataset.

Not done

A fully sparse stat_on_genes (via sparseMatrixStats) would remove the densification and the parallelism entirely, and retire the recurring parallel-chunking bug family (#71, #37, #96). That's a refactor with a new Bioconductor dependency, out of scope for a bug fix here.

https://claude.ai/code/session_01MFLoUB1VZFB5qRNYR6BxVL