Hotfix: add custom callback to avoid azureml 100 params limitation

[jfomhover]

PR into Azure/mlops-v2

We're facing a blocker when using HF's default callback to mlflow backend, that way too many parameters are sent to mlflow (more than 100). This PR creates a custom callback to avoid this issue, while the limitation is being discussed in AzureML.

Checklist

I have:

• read and followed the contributing guidelines

Changes

• removes the default mlflow callback
• implement a custom callback using only metric reporting