Metrics class should behave more like Keras's default History Callback

The `Metrics` class is a bit of a mess. It should behave more like a Keras `History` callback, and should be the returned object from a call to `saber.train()`. 

This would make hyperparm tuning much easier. E.g. a simple script could be written to collect the average performance of some metric or loss from the `Metrics` object returned by `saber.train()`, and tune against it.