add attention up/down blocks for VAE

[patil-suraj]

Our current VAE does not support having attention in encoder/decoder block as none of the pre-trained models had one. But looking at the stable-diffusion upscaling and RDM config

https://github.com/pesser/stable-diffusion/blob/693e713c3e72e72b8e2b97236eb21526217e83ad/configs/stable-diffusion/upscaling/upscale-v1-with-f16.yaml#L96

https://github.com/CompVis/latent-diffusion/blob/a506df5756472e2ebaf9078affdde2c4f1502cd4/configs/retrieval-augmented-diffusion/768x768.yaml#L62

It seems some new VAE are going to have attention in encoder/decoder. This PR adds AttnDownEncoderBlock2D and AttnUpDecoderBlock2D to suppor this.

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[patrickvonplaten]

Very nice!

One nit: Maybe let's either remove the Cross... prefix from the other attention blocks or add it here - whatever you think is best!

[patil-suraj]

I didn't put the Cross... prefix here as the model doesn't do cross-attention, there is no context/condition here. This is just self attention. So I would leave it as it is.

[anton-l]

Great use of existing modules, very clean :)