[ONNX] Add symbolic function for XSoftmax op for exporting to ONNX.

[fatcat-z]

What does this PR do?

For such customized operator, PyTorch ONNX exporter will try to call its symbolic() function for exporting to ONNX file. Add the implementation here so that it could be exported successfully.

[LysandreJik]

Might be of interest to @BigBird01

[BigBird01]

Might be of interest to @BigBird01

This is cool! Thanks!

[fatcat-z]

@BigBird01
The latest CI failure seems like is not relative to my changes. Could you please take a review and see if we can merge this change?

Thanks.

[michaelbenayoun]

Seems good to me, thanks for your great contribution!

Quick question: do you plan to add the ONNX DeBERTa export support now that you have added support for the problematic operator?

[fatcat-z]

Seems good to me, thanks for your great contribution!

Quick question: do you plan to add the ONNX DeBERTa export support now that you have added support for the problematic operator?

No, no plan for this yet.

[LysandreJik]

Yes, LGTM! Thank you @fatcat-z

[pommedeterresautee]

@fatcat-z Hi, I have a question about the way symbolic conversion is implemented. Is there any reason why you are using uint8 type for some values?
It seems to me to work with signed integers, and uint8 is not supported by TensorRT (it only supports signed integers).