Improving Conversational Recommendation Systems via Bias Analysis and Language-Model-Enhanced Data Augmentation