Towards Medical Artificial General Intelligence via Knowledge-Enhanced Multimodal Pretraining