Equitable Electronic Health Record Prediction with FAME: Fairness-Aware Multimodal Embedding