Redesign VAE for mixed-resolution tokens

Background: Diffusion models can produce occasional visual artifacts, particularly at the boundaries where different resolution features are combined during generation.

Question / Future Work: The blending of VAE-decoded low-resolution and high-resolution content near the foveation boundary can lead to noticeable color or discontinuity artifacts, suggesting a need to modify the autoencoder to support mixed-resolution tokens natively.