Robustness-Congruent Adversarial Training for Secure Machine Learning Model Updates