On the Algorithmic Bias of Aligning Large Language Models with RLHF: Preference Collapse and Matching Regularization