Direct Preference Knowledge Distillation for Large Language Models