Text this: Privacy‐Preserving Federated Multimodal Deep Learning for Sepsis Prediction Using Vision Transformers and Secure Aggregation