Convert trained models to Onnx

[George3d6]

After #473 is done

Try converting the already trained model to the onyx IR and see if that improves performance (if it does think about how this can be integrated with normal loading and saving, ideally keep both this IR and the source if certain env are incompatible with running onyx)

[paxcema]

For this, we need a flag that will constrain AutoML design space to PyTorch encoders and mixers only.

[paxcema]

Closing for now.