Are Large Language Models Good Evaluators for Abstractive Summarization?