Enhancing High-order Interaction Awareness in LLM-based Recommender Model