Add utility for training to dirty image and integrate into cross-val

[jeffjennings]

• Adds a function to train against a dirty image.
• Optionally integrates this function into a cross-validation loop.
  • Training against a dirty image is run once at the start of cross-val, with the resulting state dictionary saved. The state dict is then loaded to update each new model at the start of the training loop for each k-fold.
• Adds a test of the function.

[jeffjennings]

Great, and good point - the HD 143006 part 3 tutorial sounds like a nice place. I've added a note in #63.