Make tests less specific to a particular model

#1173 improves our test coverage by changing a lot of tests that used only the models that implement

    s ~ InverseGamma(2, 3)
    m ~ Normal(0, √s)
    1.5 ~ Normal(m, √s)
    2.0 ~ Normal(m, √s)

in various ways, to use a more general collection of models. However, many other tests remain that still have a hard-coded assumption that we are using that particular set of DEMO_MODELS. For instance, some tests assume that there is a variable called x and that it's an observation, or that there is at least one observe statement in a model. We should clean these up and make them all fully generic, making full use of the (excellent) setup in TestUtils.Models with logprior_true and varnames and other such functions used to implement the ground truth of what we expect from any given model.

Start by reviewing the changes in #1173 and the test setup in TestUtils.Models, especially logprior_true and varnames. Find remaining tests that assume DEMO_MODELS, a variable named x, or an observation, and generalize them against the model metadata. Done means the tests no longer depend on that particular model collection and the full test suite passes.