Allow for non-Keras optimizers

[matt-gardner]

Using plain tensorflow optimizers instead of Keras optimizers is beneficial in some instances (particularly when you have a large embedding matrix, as Matt Peters has discovered). We should split out the actual training loop into something configurable, to allow for different means of optimizing the same computation graph. You can still use _build_model just like we normally do, you just pull out the inputs and outputs and pass them directly into a tensorflow optimizer instead of calling model.fit().

[matt-gardner]

Note that if we do this, you could also conceivably use something like PyTorch to build your model, and just have a PyTorch optimizer. The only real benefit there would be to re-use the data processing code, though - you couldn't use any of the layers we have, or the higher-level TextTrainer API.