Narrowing Class-Wise Robustness Gaps in Adversarial Training