Disk-Based Shared KV Cache Management for Fast Inference in Multi-Instance LLM RAG Systems