Migrate LLM pipeline to custom GRPO with robust rewards

Add Cloud Arena Mathematical Model RL environment