Implementing Tests for Distributed Ops

We currently have tests to evaluate the TensorFlow version of FlowPM against our reference FastPM simulation code. They run automatically on Travis CI.
The problem now is that we want to be doing the same thing but with the Mesh implementation, which requires running the ops on a TensorFlow cluster.

We need to figure out how to run those tests automatically. The most likely answer will be to spawn a TF cluster on Travis, with like 4 CPU processes, and run the tests by connecting to this local cluster.

There are some caveats with that approach though, because as we have already seen, some ops behave differently on CPU and GPU, so...