SparseViT: Revisiting Activation Sparsity for Efficient High-Resolution Vision Transformer