Hard Adversarial Example Mining for Improving Robust Fairness