Support training variance learning (Improved DDPM) with VLB loss.

[bhosalems]

Issues has been previously discussed here, but was closed unresolved. #3287. Diffusers only support variance learning in inference, which is weird if they did not support the training in the first place. Improved DDPM has been since used in lot of other repositories like ADM and DiT.

Can this be address please.

[a-r-r-o-w]

Hi, training scripts in diffusers are not meant to be exhaustive and cover the entire literature. They are only minimal examples to get started. Learned variance denoising was supported in diffusers because released models at the time required it. Most of the time, we will prioritize only supporting the inference. Training support mostly comes from the community, and sometimes the maintainers if they have the bandwidth. We welcome any improvements and training script contributions, so please feel free to PR :)