fix self conditioning in base imagen

add self conditioning from the Bit Diffusion architecture, for base i…

…magen, elucidated imagen, and support for unet3d (video)

learned attention bias to null key / values in temporal attention

make sure temporal position generating module (depthwise conv for tim…

…e) accounts for causal vs non-causal

temporal attention needs a residual, also add the positional generati…

…ng module from vision transformers literature for time in unet3d

make null unets instantiable from config

add ability to train only super-resoluting unets, for @hmza09

allow sigma min and sigma max to be overridable at sample for elucida…

…ting ddpm

fix config because of video mode

allow for initializing with an image or video, addressing lucidrains/…

…imagen-pytorch#170