Some QOL config/saving improvements

[francoishernandez]

1. Previously, the transforms artifacts were saved only in the "root" model_path, so not accessible by default within the step_# checkpoints. This PR applies the same logic as all the other items to the transform artifacts : save the file in the step_dir, and create a symlink to the latest in the root dir.
2. To prevent validation issues, when finetuning a model with the "inference" key in its config.json, this key was popped prior to instantiating the TrainConfig. That prevents the downstream transparent use of these inference settings. This PR adds an explicit inference field to TrainConfig, which allows to properly retain it all along. It will also allow to modify the validation/inference codepath to use the "proper" settings rather than relying on defaults.