TUBench: Benchmarking Large Vision-Language Models on Trustworthiness with Unanswerable Questions