Post-Training Sparse Attention with Double Sparsity