UHGEval: Benchmarking the Hallucination of Chinese Large Language Models via Unconstrained Generation