Add an option to do the final step on a CPU

[vbuterin]

I needed to do this to be able to run the model on a 4070 (with 8 GB VRAM).

[ksasso1028]

@vbuterin a better option is to run the the model in half precision (16). using the --model-half parameter I can run the model on a 3070ti ( 8GB), but model takes less than 4 GB looking at nvidia-smi and is faster than CPU decoding.

[ghost]

@ksasso1028 --model-half seems to break inpainting though last I checked.

[ksasso1028]

@cherrerajobs whats happening is the input audio condition (for inpainting or audio init) is not being converted to half precision before running the VAE.

[ksasso1028]

I can put in a PR for this, will refer back in this comment