LocateBench: Evaluating the Locating Ability of Vision Language Models