Add epsroot-0 AdamW optimizer state for Adam-normalized attribution

Add base GPT-2 trained on wikitext 512-chunks (eps_root=0), before leave-1%-out retrains (checkpoints/step_287)

initial commit