Intriguing Equivalence Structures of the Embedding Space of Vision Transformers