[Depth Anything] How 62 Million Unlabeled Photos Created a New State-of-the-Art Vision Model