Add post-training verification

[daanelson]

I'd like to add the ability to run a prediction on a trained model and validate that the model has actually learned what we want it to. Good to verify that a trainer not only runs but also trains effectively.

I think the validation would have to live in the context of an individual training. Config design would probably be something like :

train:
  destination: <generated prediction model, e.g. andreasjansson/test-predict. leave
    blank to append '-dest' to the test model>
  destination_hardware: <hardware for the created prediction model, e.g. cpu>
  test_cases:
  - exact_string: <exact string match>
     inputs:
       <input1>: <value1>

    # shiny and new
    post-training inputs:
      <input1>:<value1>
    post-training match: 
      <output1>: <condition1>

unsure if we'd want to move training matching underneath inputs - that'd break compatibility but also be a more reasonable partition.

[andreasjansson]

Good idea @daanelson!

I think it'd make sense to put trained_model_inputs as the same level as inputs under train, so that each trained test case gets exercised.