Add test to warn for algorithmic changes

[ColCarroll]

This is a test to make sure the trace on a very simple model (literally just drawing from a normal distribution) using any of NUTS, Slice, Metropolis, or HamiltonianMC does not change at all. As a side effect, it allows for easy benchmarking of those methods between commits. I tried to comment heavily.

Interested in feedback on this test -- I've been using it on a branch I've been working on, and it has helped me avoid changing the sampling algorithm at all (though performance enhancements are still fine).

On the one hand, I worry that it will be too brittle, but on the other hand it may be a good alert if a pull request changes one of these methods.

[twiecki]

I think this is a great test to have, and the speed-comparison will be useful too.