Self-contradictory Hallucinations of Large Language Models: Evaluation, Detection and Mitigation