Enhancing Adversarial Robustness in SNNs with Sparse Gradients