Boost Vision Transformer with GPU-Friendly Sparsity and Quantization