2.0 Features to add

[ghanvert]

• Remove need to return 'loss' in 'validation_step' function in 'Trainer'.
• Parallel evaluation out of the box. #26
• Multiple evaluations in 'Trainer'.
• Multi-model support.
• HyperParameter Search integration.
• Support for all trackers:
  • MLFlow
  • Tensorboard
  • WandB
  • CometML
  • Aim
  • ClearML
  • DVCLive
• Support for multiple trackers.
• Better ways to create/define metrics.
• AcceleratorModule internal state (instead of 'status_dict', so it's easy for the user see the available options).
• Health-check support with accmt-cli. #27
• All DataCollators with a recursive approach.
• Multiple checkpoints.
• Replace handlers by callbacks.

(maybe I'll keep adding more features if I see something else missing)