Implement Stein Variational Gradient Descent

[dritchie]

http://www.cs.dartmouth.edu/~dartml/project.html?p=vgd

Only applicable in continuous spaces, but a neat idea doesn't require the definition of a variational family and should(?) be pretty easy to do in webppl with the infrastructure we have already.

[dritchie]

@stuhlmueller pointed out that we might able to extend this to models with discrete choices by alternating optimizing the continuous choices via gradient descent and optimizing the discrete ones via a derivative-free method (e.g. simulated annealing)

[null-a]

I have a first attempt at this over here. I'm not super confident it's correct, but it appears to do something sensible on this tiny model. The assumptions this makes about the model are mentioned here.