Fix onnx model export not always working properly in half precision

[lorenzomammana]

Summary

Right now when we export models in half precision using onnx as inference engine there are certain cases where the export doesn't work properly and the produced model only outputs nan values.
This pr introduce the usage of auto_convert_mixed_precision to retry exporting half precision models in mixed precision if the initial export failed

I've manually tested that finetuning, anomaly and segmentation are working properly after the changes, most of the models doesn't suffer form this issue.

Type of Change

• Bug fix (non-breaking change that solves an issue)

Checklist

Please confirm that the following tasks have been completed:

• I have tested my changes locally and they work as expected. (Please describe the tests you performed.)
• I have added unit tests for my changes, or updated existing tests if necessary.
• I have updated the documentation, if applicable.
• I have installed pre-commit and run locally for my code changes.