fix: correct model name to unsloth/gpt-oss-20b (no -instruct suffix)

fix: move unsloth import first, make wandb optional via WANDB_API_KEY

fix: install trl>=0.16 last with --upgrade to beat unsloth dep pins

fix: use CUDA devel image and pin vLLM to 0.12.0

feat: add W&B reward logging to both training scripts

feat: RFC 005 interactive rollout wrapper + multi-turn GRPO training

feat: multi-turn tool-using GRPO training

upgrade: gpt-oss-20b BF16, vLLM GRPO, Northflank env URL, max_steps=300

add training/ as real directory (Dockerfile + train.py)

add GRPO training job (Llama 3.1 8B + Unsloth + TRL)