Recurrent Context Compression: Efficiently Expanding the Context Window of LLM