Improving Inverse Folding for Peptide Design with Diversity-regularized Direct Preference Optimization