[2402.11907] Direct Large Language Model Alignment Through Self-Rewarding Contrastive Prompt Distillation