LLM Reading Tea Leaves: Automatically Evaluating Topic Models with Large Language Models