Disentangling Length from Quality in Direct Preference Optimization