Compressing Long Context for Enhancing RAG with AMR-based Concept Distillation