refactor(grpo): replace _grpo_raw_text monkey-patch with proper data structure

[abrichr]

Problem

trainer.py:228-231 monkey-patches action._grpo_raw_text = decoded onto a BenchmarkAction dataclass. This is fragile — if BenchmarkAction ever adds __slots__, the attribute silently fails and loss computation falls back to DSL reconstruction (which may differ from what the model actually generated).

Proposed Fix

Replace with a proper wrapper dataclass (e.g., GRPORolloutAction composing BenchmarkAction + raw_text) or store raw outputs in the Rollout/RolloutStep data structure.

[abrichr]

Closing: resolved by replacing the custom GRPO trainer with a minimal TRL bridge (see PR #34). The fragile _grpo_raw_text monkey-patch on BenchmarkAction dataclasses is eliminated. TRL uses standard HF datasets for storing training data, so there is no need for custom attribute injection on dataclass instances.