Cross-modal RAG: Sub-dimensional Text-to-Image Retrieval-Augmented Generation