Mlwiz

Fixed

- Bug in pre-transform and transform_train/eval not being parsed correctly

Changed

- Ray always creates a new local instance when initialized, except when working with a cluster of machines.
- DatasetInterface `__init__()` has changed to differentiate between runtime `transform_train` and `transform_eval`.
This helps when training vision models that need to perform data augmentation in the training set.
Note that the `IterableDatasetInterface` has a slightly different logic here than `DatasetInterface`.
The latter relies on two custom `SubsetTrain` and `SubsetEval` classes used by the `DataProvider`.

Changed

- Substituted `torch.load` with `dill` in some places to resolve annoying warnings
- Replaced `data_root` attributes with `storage_folder` for consistency
- Added tests for post-training utilities
- Improvements to workflow files

Fixed
-
- Fixed post-training utilities