Fast Adversarial Training against Sparse Attacks Requires Loss Smoothing