Class Optimizer

Make one optimization step. This assumes that gradients where computed previously by calling backward on a scalar variable before.

Zeros all gradients of the model parameters. This should be called after each optimization step, when the gradients are no longer needed.