Improve Gibbs dynamic model test

[mhauru]

See

Turing.jl/test/mcmc/gibbs.jl

Line 463 in 7d6a983

# TODO(mhauru) Could we do something smarter here? Maybe a dynamic model for which

The current test is a) slow, b) only compares the result against the old Gibbs sampler. Would be better to be fast and compare against an analytical result.

[mhauru]

@AoifeHughes another self-contained issue: Just requires coming up with a good model that has varying dimension (sometimes called a dynamic model), and switching one test to use that model. Dynamic models are a constant source of grief and pain for Turing.jl, but we do support them. They've caused a lot of trouble in the past for our Gibbs sampler, hence this test. The tricky part is coming up with a model that is dynamic, but for which we would have some sort of ground truth to compare against, and that hopefully wouldn't be awfully slow to sample from. I don't know what such a model would be, but ideas welcome.