Constrained or partitioned inference for Turing models. 

We currently have options to allow users to specify constraints for MAP/MLE optimisation-based learning. 


https://github.com/TuringLang/docs/pull/470

It would be interesting to also allow users to specify constraints for MCMC sampling, e.g., each chain is confined to sampling from a partition. Then, the user can merge MCMC samples from different partitions somehow. 



