Moving converters to other packages

[sethaxen]

Recently I've heard a number of reports from users that issues with Python dependencies or their installation prevent them from using ArviZ.jl (see e.g. #206). Some of these users ask if they can just use InferenceObjects.jl with Turing.jl. Since all of our converters here are pure Julia, we should prioritize moving them out of this package into either a glue package (e.g. SampleChainsInferenceObjects.jl) or directly into the package where the input to the converter is defined (as has been proposed with MCMCChains, see TuringLang/MCMCChains.jl#381).

While such users would still not be able to use our functionality that forwards to Python, they could still get the benefits of working with InferenceData, dimensional arrays, and the Tables interface.

[sethaxen]

See also StanJulia/StanSample.jl#60 for discussions on letting StanSample implement its own converter. We could then deprecate the from_cmdstan converter and point users to that.

[sethaxen]

@cscherrer it likewise would be nice to move the conversion to InferenceData from SampleChains out of this package. This could either live in SampleChains/SampleChainsDynamicHMC via (conditional) dependence or in a glue package that users would need to import. What do you think?