DuPO: Enabling Reliable LLM Self-Verification via Dual Preference Optimization