Imagine2Real: Towards Zero-shot Humanoid-Object Interaction via Video Generative Priors