cache and convert CodeFormer model

[ssube]

The CodeFormer library automatically downloads the model checkpoint from Github or Wasabi. The app should run well on a read-only filesystem with no outgoing network access (except S3), so that model needs to be downloaded and converted ahead of time.

[ssube]

The CodeFormer model has been added to the conversion script using the same code as GFPGAN (which causes some confusing log messages). That seems to work and does export an ONNX model, which is enough for the correction model menu, but the actual code is still using the CPU device/path and still downloads a checkpoint.