ReLU Strikes Back: Exploiting Activation Sparsity in Large Language Models