Local replaying of the graph

[HEmile]

Some estimators, such as https://papers.nips.cc/paper/5678-local-expectation-gradients-for-black-box-variational-inference.pdf, require local replaying of parts of the computation graph (in this case because it samples from the markov blanket, which includes the children and their parents).

Storchastic does not currently allow for such replaying. However, it does wrap around all the function calls that should make up the whole computation grap. By storing these functions and sampling steps, we could maybe recreate the computation graph for reuse in such estimators.