ELSA: Exploiting Layer-wise N:M Sparsity for Vision Transformer Acceleration