Learning from Red Teaming: Gender Bias Provocation and Mitigation in Large Language Models