Polish StableDiffusion3

[james77777778]

I'm aware that the original private implementation is missing the encoder in VAE model.
However, the encoder is necessary for image2image, inpainting and finetuning.

I'm working on bringing it back and we can expect the following features for SD3 in the future:

• image2image
• inpainting
• finetuning

Are these features planned for the KerasHub release?

TODO:

• Bring back VAE: Add VAEBackbone and use it for SD3 #1892
• Add Image2Image: Refactor MMDiT, add ImageToImage and Inpaint for SD3 #1909
• Add Inpaint: Refactor MMDiT, add ImageToImage and Inpaint for SD3 #1909
• Add T5 text encoder preset
• Add finetuning recipe for SD3