Check model by default

[torfjelde]

We now have the ability to perform some rudimentary model checks (TuringLang/DynamicPPL.jl#540), and so we should probably perform a few of these at every call to sample with an inference alg.

This PR adds these checks by default. This means that now we will give informative errors if either of the following occurs:

• Varnames are repeated, e.g. Name clash caused by submodels is hard to debug DynamicPPL.jl#427.
• Missings are found in observe statements.

We can also perform certain checks depending on the sampler, e.g. error early and informatively if the user attempts to use gradient based samplers for discrete parameters.

One question: Do we make these checks error by default or just warn the user? I'm preferential to the erroring, since these checks can always be turned off.

[torfjelde]

@yebai @devmotion

[sunxd3]

If I read it correctly, we are using https://github.com/TuringLang/DynamicPPL.jl/blob/5347acd278146449a943e3dee647d0ff9893b189/src/debug_utils.jl#L240 to track varname, then we are rely on ==.

Should we consider x[1] and x[1:2] being a clash (and even weirder subsume cases). Also, after transitioning to Accessors, we lost some generality in equality testing. For instance, x[:] concretized on x = [1, 2, 3] is !== from x[1:3] (but it was == before).

This is not a PR blocking concern, because the PR is clearly try to catch true negative.

[torfjelde]

Should we consider x[1] and x[1:2] being a clash (and even weirder subsume cases). Also, after transitioning to Accessors, we lost some generality in equality testing. For instance, x[:] concretized on x = [1, 2, 3] is !== from x[1:3] (but it was == before).

We indeed should! But yes, this is something we should deal with in DPPL, not in this PR.

But yeah, you're very correct that we should probably check subsumes rather than equality 👍

[torfjelde]

This requires #2225 as some of the fixes are present there.