ViTamin: Designing Scalable Vision Models in the Vision-Language Era