Adversarial Robustness of Supervised Sparse Coding