Boosting Adversarial Training via Fisher-Rao Norm-based Regularization