Implement chat template wrapping for prompt tokenization and add related CLI options

Fix per-head tensor pooling

fix(sub_zero): resume Stage 3 AtP per-pair + scope grad to kill OOM

Add bucket metrics and contrast delta analysis