Improving Gender Fairness of Pre-Trained Language Models without Catastrophic Forgetting