feat(training): SFT dataset + stall detection in eval rollout

fix(training): per-action reward scoring vs heuristic oracle

feat: enhance completion parsing to handle truncated JSON and `<think>` blocks

Add training notebook and benchmark runner for ChargebackOps