VimRAG: Navigating Massive Visual Context in Retrieval-Augmented Generation via Multimodal Memory Graph