HashAttention: Semantic Sparsity for Faster Inference