Training step refactor

[raehik]

this builds off of some shared work in the data step refactoring #85

• replace code obsoleted by previous data step refactoring
  • subdomain (bounding box) loading
• improve CLI
  • load forcings via either MLflow or filepath train: add --forcing-data-path CLI option to load forcing data directly instead of through MLflow #84
  • fix learning rates configuring Training CLI script prompts for unused options which MultiStepLR scheduler does not accept #90
• replace CLI
  • lower priority
• decouple internals
  • maybe not in scope. decoupling functions is one thing, but training is accomplished by building up classes and integrating with existing Torch classes in unobvious ways. the Pytorch model itself is one of the larger examples of this (see the DetectOutputSizeMixin class, used as a superclass in FullyCNN)

[mondus]

This should be going into dev

[raehik]

needs rebase (onto dev branch, which should mostly be #85 )

[raehik]

Most likely to be merged with #97 due to overlapping changes (the later steps are fairly interlinked with previous ones).

[raehik]

Re-done in #97. Closing.