DPO-Shift: Shifting the Distribution of Direct Preference Optimization