CORE-RAG: Lossless Compression for Retrieval-Augmented LLMs via Reinforcement Learning