Disentangling Visual Transformers: Patch-level Interpretability for Image Classification