feat(grpo): replace custom training loop with TRL GRPOTrainer

[abrichr]

TRL's GRPOTrainer now supports multimodal VLMs (Qwen2.5-VL). Our 793-line custom implementation duplicates advantage computation, KL divergence, reference policy management, policy gradient loss, and checkpointing.

The one genuine gap is multi-turn rollout support — TRL is single-turn. Keep GRPORolloutCollector, replace GRPO math with TRL.

References: TRL v0.20+ VLM support, WebAgent-R1.

[abrichr]

Closing: resolved by replacing the custom GRPO trainer with a minimal TRL bridge (see PR #34). This issue is the TRL migration — now implemented. The custom 793-line training loop has been replaced with TRL's GRPOTrainer, which natively handles advantage computation, KL divergence, policy gradient loss, and all the GRPO math.