Video Generation Models are General-Purpose Vision Learners huggingface.co/papers/2607.09…