feature: various improvements to dataset handling

[marcellodebernardi]

This PR is a depressingly confusing mish-mash of several bugfixes, improvements, and clean up. The main changes are:

1. The DatasetGenerator has been simplified, cleaned up, made properly async, and can now handle column-wise data augmentation (earlier only row-wise).
2. The traces of the smolagents agents are now logged to mlflow for easy tracking
3. Improvements to mlflow tracking, chain of thought is more informative, etc
4. Include the EDA report in the model bundle and mlflow for convenience
5. Miscelanneous bugfixes