Add SKIP_EVAL flag to sft_on_hf.py for faster training-only runs

Algorithm Distillation: grader v2 with belief_accuracy + SFT pipeline