Sparse then Prune: Toward Efficient Vision Transformers