Learning Missing Modal Electronic Health Records with Unified Multi-modal Data Embedding and Modality-Aware Attention