Peeling the Onion: Hierarchical Reduction of Data Redundancy for Efficient Vision Transformer Training