Can Large Language Models be Good Emotional Supporter? Mitigating Preference Bias on Emotional Support Conversation