Impl SaveToNpz for Optimizers

[jafioti]

On long training runs it's required to checkpoint models in case something goes wrong part way through. When only saving the model, and resetting the optimizer state, training takes much longer to resume. It would make sense therefore to save and load the optimizer as well. I believe this can be implemented with the same npz architecture as is currently implemented for TensorCollection modules.

[coreylowman]

Related to #101

[nkoppel]

I think I have an approach that will work well for this:

1. Allow converting Gradients to and from Models to make serializing them easier.
2. Create new optimizer objects that store their Gradients as Models.
3. Serialize these objects instead of the optimizers themselves.
4. Remove TensorCollection subtrait of serialization traits so that we can create a custom implementation for optimizers.