Expand model correctness tests

[null-a]

• Check that mu is computed as the correct function of latents & data? (Add tests to check correctness of model's mu computation #52)
• Check that the response distribution has the correct parameters? (test_expected_response_codegen might already do this to some degree. Perhaps that can be reworked along the lines of test_mu_correctness -- using .fitted('expectation') to generate actual values, and comparing against the output of the mean method of a pyro distribution.)
• Extend code gen tests to check that the response is observed, and that it comes from the expected family?

[null-a]

Check that mu is computed as the correct function of latents & data?

Here's one way we might do this. We might define test cases that contain something like the following:

df = pd.DataFrame({
    'y': [0., 0.],
    'a': pd.Categorical(['a0', 'a1']),
    'b': pd.Categorical(['b0', 'b1']),
})
model = defm('y ~ 1 | a:b', df)
def expected(df, coef):
    return (((df['a']=='a0') & (df['b']=='b0')) * coef('r_a:b[a0_b0,intercept]') +
            ((df['a']=='a1') & (df['b']=='b1')) * coef('r_a:b[a1_b1,intercept]'))

... which would allow us to check that generated models correctly compute the location parameter of the response distribution with something like:

fit = model.generate(backend=numpyro).prior(num_samples=1)
actual_mu = fitted(fit, what='linear')[0]
expected_mu = expected(df, partial(get_scalar_param, fit)).to_numpy()
print(np.all(np.equal(actual_mu, expected_mu)))

I like this because such tests are essentially deterministic, and they're easy to write. While it wouldn't guarantee that generated models have the correct semantics, it would give us confidence that all backends compute mu in the same way, and provide reassurance when making changes to e.g. code generation. (e.g. #10.)

Eventually we might even consider generating the expected functions from the model definition itself. I guess this would be of most interest if we were also generating model descriptions in statistical notation (#33). If these shared a common implementation (you'd need generate something like the expected function when generating the math description) then these tests would help convince us that the math and code we generate are consistent.