Switching to the new lightning trainer

[dcfidalgo]

This is an issue to keep track of the process of switching to the new lightning trainer:

• Basic training with lightning trainer
• Configure default logging (wandb, tensorboard, csv)
• Configure default model checkpointing, loading best weights in the end Allow loading model via LightningModule.load_from_checkpoint #541 Feat lightining trainer: Load best weights + save pipeline to output dir #542
• Save trained pipeline to an output dir after training Feat lightining trainer: Load best weights + save pipeline to output dir #542
• Add default early stopping Feat: add early stopping by default to lightning trainer #544
• Use GPU automatically if available LightningTrainer: Train on gpu by default if available #545
• Add default learning rate scheduler Adds parameter for lr schedulers for the lightning trainer #546
• Adapt TuneExperiment to use new trainer Feat: tune experiment with new lightning trainer #548
• Add best_metrics Feat/log best metrics for the new trainer #550
• Adapt documentation/tutorials Feat: Adapt docs to reflect new trainer #551

[dcfidalgo]

I'm waiting for the 1.3.0 release of Pytorch Lightning and then i will make a release that supports both trainers (AllenNLP's and PL's). After this release i will remove the AllenNLP Trainer and bumb up the AllenNLP version.