LongRAG: Enhancing Retrieval-Augmented Generation with Long-context LLMs