Try out various U-Net architectures

[jacobbieker]

Use Pixel Shuffle and other ones instead of the standard transpose convolutions, self-attention in the lowest downsample layer, add upsample with blur