feat: finalize swarm tooling and submission artifacts

feat: add rollout observability — JSONL logs + per-step action/reward printing

fix: load 4-bit model manually before passing to GRPOTrainer

fix: use QLoRA (4-bit + LoRA) for Qwen3.5-9B training

fix: add LoRA + gradient checkpointing to fit 9B on single H100

fix: drop vLLM colocate (version conflict), use native GRPO generation

feat: rewrite training to use TRL rollout_func + OpenEnv multi-turn pattern

feat: add --mode infer for Qwen inference test, default to Qwen3.5-9B

feat: fix openenv 0.2.1 API, add deployment files and GRPO training

feat: add openenv wrapper and training stub

chore: add project scaffolding