Fairness Overfitting in Machine Learning: An Information-Theoretic Perspective