Avoid whoami rate-limit during job submission

Harden HF job token wiring and persist full training outputs

Fix GRPO batch/generation mismatch: auto-adjust num_generations; set launcher default to 2.

Fix HF Jobs bootstrap (pin transformers/trl, drop torchao stack); add reward and trainer JSONL logging; stabilize launch_job.

Fix: Mock vllm and llm_blender to stabilize GRPOTrainer in HF Jobs environment