support for CLIP guided latent diffusion

[dmarx]

should be able to adopt a lot from the VQGANImageModel , really all that's different is how candidate frames are generated. If that's not already reasonably abstracted, it needs to be.