docs(modeling): add cluster-free (non-SGE) run instructions

[cbiar]

Makes scripts/modeling/'s README sufficient to reproduce the pipeline without an SGE cluster.

I verified the gap by taking a clean checkout of exactly what ships (code + raw data, no processed CSVs, no models) and following the README in the pinned env. Findings:

• The submit_*.sh wrappers are ordinary bash scripts — the #$ lines are SGE directives that plain bash ignores. Confirmed bash scripts/submit_preprocess.sh runs end-to-end with no scheduler.
• Following the old Option B literally ran without errors but produced silently incomplete results: only the full subset trained, and merge.py loaded "0 holdout evaluations" (so combined_df's holdout columns were empty).
• The step-1 comment claimed processed CSVs were "already provided" — stale, since they're regenerated by preprocess.py.

Change (docs only, one file)

Rewrites the "without a cluster" section into:

• B1 (recommended): run the wrappers with bash scripts/submit_*.sh for full published-run coverage (all subsets, every best model).
• B2: granular per-stage python calls, now with the explicit per-subset input to output mapping and a note that every best model must be evaluated to populate combined_df.
• Corrects the "already provided" claim.

No code or data changes.

🤖 Generated with Claude Code