Removing Device generic from Gradients & optimizers

[coreylowman]

This is a refactor of the Gradients object to remove the need for the existing generic argument on it. The presence of it meant that you had to specify the device on Optimizers, since they heavily use Gradients as a store of extra buffers.

Turns out, since gradients stores Box<Any>, it can just depend on getting the device from the type parameter passed into its methods, like it does with Shape & Dtype.