MARS: Learning Modality-Agnostic Representation for Scalable Cross-Media Retrieval