feat(grpo): add checkpoint resume

[abrichr]

train() always starts from step 0. With 1000 steps at ~6 min/step, a crash means starting from scratch.

Add resume_from_checkpoint that loads:

• LoRA weights
• Optimizer state
• Step counter
• Log entries

[abrichr]

Closing: resolved by replacing the custom GRPO trainer with a minimal TRL bridge (see PR #34). TRL natively supports resume_from_checkpoint, which handles loading LoRA weights, optimizer state, step counter, and training logs.