Add onnx support

[davanstrien]

Currently, onnx isn't used for inference of transformer models. Alongside using ray one approach to improve inference throughput again would be to:

• download models
• do the onnx export from the model
• load using onnxrunntime
• catch errors for unsupported models
• do some initial validation of results?

Hopefully, we get some of this for free with Optimum if we wait a bit: https://github.com/huggingface/optimum