fix(train): v2 — fix collapse root causes + best-ckpt tracking + GGUF export

fix(train): also clone log_probs to escape inference_mode context

fix(train): clone prompt/completion tensors after inference_mode generation

fix(train): tokenize-free log-prob recompute + switch default model to Llama 3.1

fix(train): allow gradients through compute_log_probs for current policy

fix: Qwen3-8B-Instruct support — model name, think token stripping

feat: hackathon improvements — benchmark dashboard, GRPO training pipeline, grader fixes