perf(grpo): use PEFT named adapters for reference policy

[abrichr]

_compute_ref_log_probs clones all LoRA weights, swaps to reference, runs forward, restores — per step per rollout (120+ swaps per training step).

Use PEFT's set_adapter() or named adapters ("current" / "reference") instead.

Also: LoRA state capture uses fragile "lora" string matching on state_dict(); use get_adapter_state_dict() if available.

[abrichr]

Closing: resolved by replacing the custom GRPO trainer with a minimal TRL bridge (see PR #34). With beta=0.0 (no KL penalty), there is no reference model needed at all — eliminating the fragile LoRA weight clone/swap/restore cycle. TRL also handles reference policy management natively via PEFT integration for cases where beta > 0.