GIEBench: Towards Holistic Evaluation of Group Identity-based Empathy for Large Language Models