Nengo version of ModelCheckpoint callback

[bmorcos]

Right now it's possible to use tf.keras.callbacks.ModelCheckpoint to save (keras) model weights during training under NengoDL, and then use sim.keras_model.load_weights(...) afterwards. However it might be nice to have our own version of this that uses the native save_params and load_params directly.

I briefly looked at the ModelCheckpoint code and after a quick scan it looks like it should be pretty reasonable to subclass.