Multimodal Relation Extraction with Cross-Modal Retrieval and Synthesis