feat: add episode trace, refresh training dataset, and update eval metrics

feat: add structured pruning action and random baseline policy

refactor: harden imports, add training extras, and rewrite README