Flow to the Mode: Mode-Seeking Diffusion Autoencoders for State-of-the-Art Image Tokenization — arXiv2