feature(ai): enable GAN upscalers with spandrel

[ad-astra-video]

New addition to stack is adding adding chaiNNer/spandrel to enable adding multiple models.

To ease the downloading of the different checkpoints I made a huggingface repo to bring them in and link to the source repo of the models: https://huggingface.co/ad-astra-video/upscalers. Adding the files to a Huggingface repo also adds some security precautions run on the files when they are uploaded to lessen chance a malicious file is downloaded.

• Upscale multiple is set by the model and not modified so not change to API needed at this time.
• Upscale multiple is limited based on size of input. With testing on a A6000, 1024x1024 4x upscale takes about 22GB vram and a 512x512 4x upscale takes 9-12GB vram. There is some variation between models but not significant.

Thoughts on limiting image inputs to 512x512 for upscaling to enable better coverage with 16gb gpu requirement?