How to Train Your DRAGON: Diverse Augmentation Towards Generalizable Dense Retrieval