Refactor training logic

[billylanchantin]

Adds:

• %Training.State{status: :cont | :halt}

This way the fun on %Training.Callback{} can have the spec:

• fun :: State.t() -> State.t()

instead of wrapping the output in a tuple. It also means we can write code like this:

state =
  state
  |> run_callbacks(callbacks, :before_training)
  |> run_training(callbacks, dmat, objective)
  |> run_callbacks(callbacks, :after_training)

since now run_callbacks/3 and run_training/4 (new) can branch on state.status.

I did some other code shuffling while I was in there, but none of it is crucial.