Add an example of exporting BartModel + BeamSearch to ONNX module.

[fatcat-z]

This PR is aimed to deliver an example to describe how to export BartModel + BeamSearch to ONNX model and inference it by ONNXRuntime.

These new example files were added without updates to existing Huggingface model code. It mainly focuses on:

1. Providing an entry file run_onnx_exporter.py to run this example.
2. In generation_onnx.py file, adding a BARTGenerator model to wrap the native BartModel code. To support Beam Search, we need to convert the whole model to a TorchScript module finally. This new BARTGenerator model will trace the Encoder and Decoder of BartModel to make them be compatible with torch.jit.script() method.
3. In generation_onnx.py file, copying Beam Search code from generation_utils.py and updating these code so that beam_search function could be converted to TorchScript by calling torch.jit.script method.

[patrickvonplaten]

Think we just need to run make style once and then we can merge - lgtm!

[NielsRogge]

Thanks for adding this! Could you perhaps add a README to this example, to explain what the example is about and how to use it?

[garymm]

Thanks for this @fatcat-z.
Could you please send another PR that adds a README and addresses my comments?

[NielsRogge]

Hey @fatcat-z,

kindly pinging you for taking into account the comments on this PR (perhaps in a follow-up PR), if possible!

[fatcat-z]

Thanks for adding this! Could you perhaps add a README to this example, to explain what the example is about and how to use it?

I'm working on a new PR to address your and Gary's comments.