RULE: Reliable Multimodal RAG for Factuality in Medical Vision Language Models