Refactor: compute metrics in experiments class

[jxmorris12]

• Compute metrics in experiments class
  • Add PyTorch functions for metrics in metrics.py (that way metrics can run on GPU not CPU)
  • add metrics for Finetuning (todo: add metrics for Retrofitting)
  • remove logic for metric computations in train.py
• Support logging metrics multiple times per epoch via --logs_per_epoch arg
• Add loggers everywhere (instead of simple print statements); log to both file & stdout
• Add end-to-end tests for training
• Add TensorRunningAverages object that can average metrics that are logged multiple times, so we can log the average to w&b instead of logging them many times. (Also adds associated file utils.py)

[jscuds]

Closed because I merged exp-metrics into rf-loss