Convert training model to pytorch-lightning

[faroit]

lightning has matured enough to be used to refactor open-unmix.

[faroit]

lightning would significantly clean up the training code but when using learning rate scheduling and distributed training code from lightning, users won't be able to reproduce the current pre-trained models.

Since results are reported in publications, we want to reduce confusion, and therefore do not currently plan to retrain the model without major core changes. Hence, I propose to skip this feature until we have a core change in the model