fix(grpo): align prompt format across SFT, GRPO, and CoT warmup

[abrichr]

Problem

Three different prompt formats exist in the codebase creating distribution shift:

• SFT (next_action.py): includes action history, "Thought: ..." format, step counter
• GRPO (trainer.py _build_agent_messages): "Goal: ..." with no history or thought
• CoT warmup (cot_warmup.py): "Instruction: ..." prefix, different history format

The model trained on SFT encounters different prompts during GRPO, degrading early GRPO performance.

Proposed Fix

All three should use a shared prompt builder with configurable components (history, thought, step counter). This ensures the distribution seen during SFT matches what the model encounters during GRPO rollouts.

[abrichr]

Closing: resolved by replacing the custom GRPO trainer with a minimal TRL bridge (see PR #34). The three divergent prompt formats (SFT, GRPO, CoT warmup) in the custom implementation are eliminated. TRL uses a unified data pipeline with standard HF dataset formatting, ensuring consistent prompt construction across training stages.