Enhancing Model Robustness and Fairness with Causality: A Regularization Approach