Add config-driven pipeline CLI; make the example configs executable

[DaoyuanLi2816]

Closes the gap where examples/configs/*.yaml had no consumer in the codebase:

• python -m labelbank.run --cfg <yaml> — stage: retriever runs the full loop the README describes (zero-shot eval → rank the whole bank → gold_first pool mining → contrastive retrain → re-evaluate, for mining_rounds rounds), saving the adapter, per-split rankings.parquet and metrics.json. stage: reranker consumes that parquet and trains the listwise reranker (completion-only, gold position shuffled). Rounds continue one adapter; the docstring notes the competition's fresh-per-round protocol (measured separately in the upcoming experiment example).
• Both example configs rewritten to the flat schema the CLI actually parses (reproduce_competition.yaml keeps the original v0 hyperparameters incl. pool 50, NF4, the Eedi instruct prefix, fold-0 eval); a test asserts the shipped configs load.
• 10 new CPU tests: config defaults/unknown-key rejection/stage validation, generic-format column remapping, seeded holdout and fold splits. Suite: 73 passed.
• README: install matrix now lists the existing [rerank] extra; pipeline section shows the two CLI one-liners.
• competition/README.md: public-LB typo in the intro fixed (0.54 → 0.56, matching the two other mentions in the same write-up).

🤖 Generated with Claude Code