Add a Turing integration test that isn't running the entire test suite

[penelopeysm]

Technically, the error in #207 did show up in the CI runs of #205, but given that that CI job is running the entire Turing test suite which can fail on the best of days because a giraffe halfway across the world burped, I suppose it is easy to overlook it. It's also quite wasteful because most of the Turing tests have nothing to do with Libtask. I think we can do better by just having a much smaller 'test suite' of

using Turing

@model function f()
    x ~ Normal()
    2.0 ~ Normal(x)
end

sample(f(), SMC(), 10)
sample(f(), PG(10), 10)

I suppose we could do correctness checks as well, if desired (bump the sample counts and check that mean of x is around 1.0). And I guess we could add a basic PG-in-Gibbs test that's similar to the above.

Realistically I don't think that anything else needs to be tested. Numerical accuracy checks for complicated models aren't really Libtask's problem, that would be AdvancedPS / Turing's problem.