Boosting Vision Semantic Density with Anatomy Normality Modeling for Medical Vision-language Pre-training