Improvements to TestUtils (follow-up from #360)

[torfjelde]

When working on #360 I realized there were quite a few features we're not covering with the models provided in DynamicPPL.TestUtils, e.g. link and invlink are not covered since the current models do not include any constrained variables. This PR adds some additional methods in DynamicPPL.TestUtils which ought to be implemented for the models and a more general test_sampler_on_models.

EDIT: No need to review this until we've merged #360 but I wanted some of these tests for SimpleVarInfo to make sure it works properly.

[torfjelde]

@devmotion would you have a look at this so we can just merge this together with #360 ? This PR only makes changes to testing, and it would be nice to make sure that #360 passes with the additional tests provided in this PR before making a release anyways.

[devmotion]

Just

Suggested change

	test_sampler_on_models(meanfunction, models, sampler, args...; kwargs...)
	test_sampler(meanfunction, models, sampler, args...; kwargs...)

?

[devmotion]

Some other points:

• Since we test the sampler but not the models maybe that should be the second argument?
• Is it useful to be able to specify a meanfunction? Wouldn't it be sufficient if we provide defaults for Chain, Vector{<:NamedTuple}, Array{T,3} (or whatever common types are) as marginal_means(chain). We would always use marginal_means but users could still implement marginal_means for custom types.

[torfjelde]

Since we test the sampler but not the models maybe that should be the second argument?

I'm happy to make that change:) The reason why I did the current ordering is because it's similar to sample which has model first.

Is it useful to be able to specify a meanfunction? Wouldn't it be sufficient if we provide defaults for Chain, Vector{<:NamedTuple}, Array{T,3} (or whatever common types are) as marginal_means(chain). We would always use marginal_means but users could still implement marginal_means for custom types.

With the number of different Transition types we have lying around, I think it might be worth just allowing someone to quickly do

test_sampler(...) do chain, vn
    # Compute whatever.
    ...
end

But happy to make it a separate function that can be overloaded or something too 👍

[torfjelde]

This means that currently on #master anything involving NoDist + linked VarInfo is producing incorrect results (which, AFIAK, we were unaware of) 🙃

[yebai]

Anything missing to merge this PR into #360?

[torfjelde]

The only thing remaining is to resolve the current open comments in this PR, but these are very minor.

I would recommend that we get #360 approved before we merge the two though, and then once that's approved we merge this into #360 . Then we leave it up to bors. This will reduce the amount of code you'll have to look at in one go when reviewing the PRs.

[torfjelde]

I've made the requested changes here; should be good to go for merge once #360 has been approved.

[devmotion]

Looks good (didn't recheck everything in detail).