Reinforcement Learning from Multi-role Debates as Feedback for Bias Mitigation in LLMs