Compile BUGS models to a Distributions.jl Distribution type

[yebai]

Another very useful target for BUGS models is a Distributions.Distribution object. This would allow us to use these models as a new Distribution in other PPLs including Turing.

Here are two concrete use cases:

1. use BUGS language to specify a complicated hierarchical prior, then the DAG would allow us to perform re-parameterisation optimisation etc. See e.g. here
2. use BUGS language to specify an HMM/SSM model, then DAG would allow us to perform SMC/PMCMC inference efficiently

[sunxd3]

For now, we will postpone supporting this util JuliaStats/Distributions.jl#1762 or similar interface is implemeted.

[sunxd3]

Update on this JuliaStats/Distributions.jl#1803

[sethaxen]

@sunxd3 what's your usecase for a NamedTupleVariate?

[sunxd3]

@sethaxen mostly an interface nicety, I am thinking compiling BUGS to a NamedTupleVariat, then the model can be used in a Turing model (and vice versa).