Refactoring AD Tests

[willtebbutt]

Currently, there are many situation in which we run an entire sampler in order to test AD, such as many cases in https://github.com/TuringLang/Turing.jl/blob/master/test/mcmc/abstractmcmc.jl .

This is quite compute intensive. @yebai pointed out that perhaps we should just check that the gradient runs and is correct, rather than thousants of mcmc chains.

[yebai]

We should probably implement some functionality to benchmark various AD backends during this refactoring.