Interior interpretability with attention rollout: contraction and propagation profiles in Transformers