feat: multi-model and multi-label classify-warc output

[malteos]

Summary

• classify-warc now scores each record against N fasttext models × M labels per model in one pass, via parallel-list --model-repo / --model-file / --labels flags (with * = all labels of that model, the default).
• Output column naming: score_<label> for a single model, score_m<idx>_<label> for multiple — so two models can share a label name (e.g. both GneissWeb classifiers emit __label__cc) without collision.
• --overwrite to replace an existing output instead of failing fast.
• --resume-from-output validates that the prior CSV's header byte-matches the new run's schema, with a structured added/removed-columns diff on mismatch.
• Per-label score stats land in the sidecar as score.<column>.<stat>; the sidecar's arg.score_columns row records the resolved label vocabulary so a run is reproducible from the summary alone.
• Notebook parameterized on SCORE_COL (auto-detected from score_* columns, with a prediction_score fallback for legacy CSVs).

Test plan

• make check — ruff lint, ruff format, pytest (63 pass, 1 skipped real-model)
• Real-model end-to-end (pytest --run-real) — language-id classifier round-trip with new column naming
• Manual: science model alone, default labels → header has score___label__science and score___label__cc, per-row sum ≈ 1.0, quantum-mechanics doc → 1.0 science
• Manual: sci + quality together with both defaults — produces score_m0___label__cc, score_m0___label__science, score_m1___label__hq, score_m1___label__cc (auto-namespaced, no collision)
• Manual: --resume-from-output against a CSV with different columns errors with a structured added/removed diff
• Manual: --overwrite warns + rewrites; absence still errors with a message pointing at the new flag
• Notebook executes end-to-end against existing legacy CSVs via the prediction_score fallback path