Add --curated-dir to soma reproduce to reuse a pre-curated manifest

[clemsgrs]

What

Adds --curated-dir DIR to soma reproduce as a third manifest source alongside --raw-root (curate from raw) and --from-run-dir (re-score an existing run). When given, it points the pipeline straight at an existing dataset.csv + splits.csv and skips curation entirely.

Why

Curation is a deterministic raw -> manifest step that can cost minutes — the HEST-bench IDC task explodes ~35k spots to lossless PNGs — yet soma reproduce re-ran it on every invocation. This bit during the manual hest/IDC uni2 reproduction, where the same validated manifest was re-derived on each run. --curated-dir lets you curate once and reuse across seeds/encoders.

Design

Chosen over silently reusing an already-populated --out-dir: an explicit flag keeps the "this manifest is good, don't re-derive it" assertion with the caller and never picks up a stale manifest without a signal. The flag is:

• dataset-type-agnostic — the spatial_expression targets.npy / genes.json sidecars are resolved by the Manifest loader relative to dataset.csv, so it needs no per-benchmark knowledge;
• family-aware — nests per-member under fan-out (eva, hest);
• fail-fast — errors clearly when the dir is not a manifest;
• deferential to --from-run-dir when both are given.

Tests

• tests/test_benchmarks_cli.py: +3 (skip-curation, missing-manifest error, three-mode error message)
• docs/cli.rst regenerated from build_cli_rst(); docs-sync + Sphinx build green
• Full test_benchmarks_cli.py + test_docs.py: 41 passed