[rich-iannone]

With this PR ablations can now be exported for training elsewhere and read back with export() / load_results(), which check a manifest of data and plan fingerprints. Results can be cached with cache= / evaluate_version=, and runs can be parallel with run(max_workers=, executor="thread" | "process").

New strategies and checks: drop_all_but_each() (what each step adds on top of the others, with a redundant verdict), drop_groups(), a pb.ablation.Stratified control, check_determinism(), and annotate(), which writes the results into the validation's step notes. Ablations can also be defined in a YAML ablation: block and built with pb.yaml_ablation(), with evaluate: behind allow_python. A new pointblank[ml] extra provides pb.ablation.sklearn_evaluator() for a ready-made scikit-learn evaluate function. Tests run across Polars, pandas, DuckDB, SQLite, and PySpark, and the data ablation user guide has new sections for each feature.