fix(grpo): fresh reward design — runbook-compliance scoring

fix(grpo): eliminate zero-advantage collapse from stale rewards

Merge remote-tracking branch 'hf/main' into train/v1

fix(training): fix GRPO reward bugs + add v3 notebook with SFT warmup

fix(reward): harsh -1.0 for zero JSON output, -0.30 for zero task completion

fix(rollout): safe model unwrapping for accelerate/PEFT/DDP in _generate_no_vllm

fix(rollout): replace generate_rollout_completions (vLLM-only) with direct model.generate()

fix(training): reward signal flow + new v2 notebook

modifying the training script to be more memory efficient: Unsloth, also reduced the number of steps to 15 instead of 40

writing the GRPO script to train qwen1.7 on A100 GRPO