JAXNS distributed over cluster (using ZMQ)

[Joshuaalbert]

1. We setup the model as a pickle-able structure, and the inputs to it as well.
2. A load-balanced fabric using ZMQ enables distributing likelihood calculations out.
3. The central NS algorithm is dissected, so that likelihood evals can be outsourced.
4. Likelihoods are still jittable onto target hardware, and AD-able.