feat(attack-ui): add ground-truth aware evaluation and sample GT auto-selection

feat(metrics): add rollout-based attention shift metrics (W1 and JSD)

feat(apuana): add local-only cluster tab flow, history diagnostics, and metric selection updates

fix: random_start=False em PGD para reprodutibilidade

git commit -m "feat: expand ViTViz with experiment pipeline, metrics module, model loading improvements and report UX" -m "- add experiment infrastructure with configs, sweep scripts, notebooks and Makefile tasks