Fix cross validation for classification models and update pytorch-lightning

[melanibe]

• Various fixes to metrics aggregation/download/paths for classification cross-validation runs:
  ** Fixed metrics aggregation for multi-gpu
  ** Fixed path for test metrics
  ** Added split recording for test set inference during CV
  ** Removed Ensemble from the folds taken into account for aggregated results in classification CV
  ** Removed Val inference for Ensemble (as val of parent fold is in training set of ensemble)
  ** Make CrossValidation folder follow a simple Train/ Val/ Test/ structure
• Add end-2-end test for classification model in PR build
• Updating pytorch-lightning to 1.2.7
• Fixed ddp setting when multi-node but only 1 gpu per node
• WARNING: removing test for training multi-node as new PL version breaks multi-node training.
  Closes No end-to-end AML job for classification model #433

[dccastro]

Looks great to me! Just proposed very superficial changes.

[dccastro]

Awesome, all clear 👍