Faster reassign-cells scoring and pb tree folds; give back preload reservations

[YPARK]

Two speedups to the pseudobulk steps, with identical results, plus a fix that gives back preload memory.

Faster DC-Poisson scoring (alg/dc_poisson.rs)

DcPoissonStats::gene_sum / log_gene are now stored gene-major (g*K + k, with a new idx(k, g) helper). They used to be group-major. compute_log_probs_restricted scores all allowed groups in one pass over the entity's nonzero genes, reading K values next to each other for each gene. It skips the K×M table stride it used to take for every group.

Each group's sum still adds the same terms in the same order (f32 cache → f64, in gene order). The leave-one-out score for the current group is computed exactly as before. So scores are bit-for-bit identical. A new test checks this with to_bits() against the per-group reference, on full and partial candidate sets, after moves. Partial candidate lists are scored by their position in the list, using a scratch buffer for each thread. If a list repeats a group, each copy gets the same correct score, just as in the old per-group scoring. A test covers this.

Public API: the fields are still pub, but their memory layout has changed. Nothing outside this file reads them: I grepped data-beans, senna-rs and pinto-rs. pinto only uses the refine_* / proposer API, which has not changed.

pb tree folds (collapse_data/pb_tree.rs)

NodeProfiles::new, residual_variance and side_sums fold into dense accumulators that cover every gene. Rayon used to split small leaves into many tasks of a few cells each, and each task zeroed and merged 36k-long vectors. Now each task takes at least 256 cells (with_min_len). The math is unchanged. Rayon's fold grouping was already decided at run time, so this does not make the output any less deterministic.

Preload reservation (sparse_io/helpers.rs, zarr + hdf5 backends)

• New reserve_preload(nnz, what) -> Option<PreloadReservation>. The returned RAII guard is stored next to the preloaded arrays. It gives back its share on clean_preloaded_* and when the backend is dropped. A clone reserves its share again.
• preload_columns / preload_rows now do nothing if the data is already preloaded.
• preload_within_budget stays for compatibility (it keeps its reservation, as before). preload_reserved_bytes() reports the bytes reserved now.
• Tests: guard release and refusal (unit), and a backend test that runs preload → repeat → clean → drop on zarr and hdf5.

Measured

Data: 10k_BMMNC (8,195 cells × 36,591 genes, 128 groups, 16 threads). Each figure is the wall time between log lines, from 3 runs per binary of senna topic --pb-tree refined:

step	before	after
reassign cells (includes its prep)	7.2 / 7.2 / 9.2 s	3.4 / 3.5 / 4.2 s
pb tree	8.1 / 8.1 / 7.9 s	1.4 / 2.1 / 1.6 s

Identical results: reassign cells: 14438 moves in every run. pb_tree.json, cell_to_pb.parquet, pb_gene.parquet, coarsening.json and pb_reference.json have the same md5 before and after.

Checks: cargo test --features sim passes. Clippy reports no lints in the changed files. The -D warnings errors it shows are new rust 1.98 lints in files this PR does not touch, and are already present on main.