improve low resolution aug noise level conditioning (make sure attent…

…ion can attend to ddpm time and aug level separately)

allow for prediction of x0, due to upscaling unets potentially using …

…that objective, discovered in appendix pseudocode by @marunine

add ability to condition all batch elements on single augmentation no…

…ise level for low resolution conditioning

add p2 loss reweighting for discrete time gaussian diffusion training

make sure classifier free guidance is applied to mean pooled projecte…

…d text hidden conditioning to time

add ability to turn on linear cross attention, to further strengthen …

…conditioning, if needed

fix linear attention

fix linear attention

extra layernorm for non-attention route of text conditioning

additional non-attention based text conditioning along with time