Allow MeanField constraints

[wouterwln]

This PR allows the following syntax:

@constraints begin
    q(x, y) = MeanFieldI()
    q(z) = MeanField()
end

While building this, I also found a bug. When mat is a matrix valued RV and the user specified:

@constraints begin
    q(mat) = q(mat[begin])..q(mat[end])
end

and obscure error was thrown. Because of the way we encode ranges in the current implementation we actually don't allow multidimensional splitted ranges so a helpful error message will now be thrown.

[bvdmitri]

Is at least

q(mat) = MeanField()

allowed?

[wouterwln]

@bvdmitri of course it is, it goes through a different channel since MeanField is treated differently

[bvdmitri]

Good! Can we add a testcase where we apply MeanField for a matrix-based or even higher dimensional variables just in case? Such that we don't break it accidentally, doesn't need to be a valid model, just usage of matrix-based variables.

[codecov]

Codecov Report

Attention: Patch coverage is 93.75000% with 1 lines in your changes are missing coverage. Please review.

Project coverage is 92.05%. Comparing base (1ded935) to head (8b6b0ac).
Report is 10 commits behind head on dev-4.0.0.

Files	Patch %	Lines
...onal_constraints/variational_constraints_engine.jl	92.85%	1 Missing ⚠️

Additional details and impacted files

@@              Coverage Diff              @@
##           dev-4.0.0     #189      +/-   ##
=============================================
+ Coverage      91.92%   92.05%   +0.13%     
=============================================
  Files             15       15              
  Lines           1845     1863      +18     
=============================================
+ Hits            1696     1715      +19     
+ Misses           149      148       -1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[wouterwln]

Added a test that tests matrices

[bvdmitri]

I like it!