Chain-of-Specificity: An Iteratively Refining Method for Eliciting Knowledge from Large Language Models