VulDetectBench: Evaluating the Deep Capability of Vulnerability Detection with Large Language Models