feat(cluster): species-parameterize CORUM + MSigDB benchmark generation (#238)

[mat10d]

Summary

Cluster benchmark generation (workflow/lib/cluster/scrape_benchmarks.py) hardwired human sources, so mouse (and other non-human) screens could not build the CORUM/KEGG group validators. Closes #238.

• CORUM (get_corum_data / generate_corum_group_benchmark): now accept species_id. Human keeps its native file_id=human export unchanged (zero regression); other species are pulled from the CORUM complete export and filtered by the organism column (CORUM has no per-organism download for them).
• MSigDB (generate_msigdb_group_benchmark): now accepts species_id and resolves the collection per species. Human = KEGG MEDICUS (unchanged). Mouse has no native KEGG collection in MSigDB, so the native mouse Reactome collection is used. An explicit url still overrides.
• STRING and UniProt generators were already species_id-aware — no change; the companion analysis PR just starts passing them the species.

Note on the mouse "KEGG" slot

For mouse, the second group validator is native Reactome data occupying the slot named kegg_group_benchmark. This was a deliberate choice (native mouse pathways, no orthology inference) over orthology-converting human KEGG. Functionally it is a valid pathway group benchmark; the name is retained to avoid churn in config keys and downstream labels.

Verification

Live smoke test (not covered by small_test_analysis, which reads static committed TSVs and never invokes these generators):

• CORUM human path byte-identical request → 5,628 complexes (unchanged).
• CORUM mouse → 1,333 complexes → 3,893 rows / 1,322 groups (matches an independent native-mouse rebuild).
• MSigDB mouse → native Reactome groups.
• ruff check + ruff format --check clean.