Reimplement stochastic differentiators with C++ and new interface

[zaqqwerty]

There are a couple issues with the old SGD:

1. Too slow compared to the other differentiators
2. Interface is cluttered

To resolve these concerns, we need to split these differentiators into separate classes for each kind of stochasticity, and the ops need to be pushed into C++ implementations as was done for the parameter shift diff. @jaeyoo see PR #372 .