Parallelise MCMC kernel exploration across tempered densities / chains

[HarrisonWilde]

During a "macro" step of the PT process, multiple samplers independently take >=1 "micro" steps (in accordance with swap_every passed to tempered or tempered_sample) within each tempered chain. The samplers all carry out these "micro" steps independently from all other samplers / chains, meaning we should be able to parallelise across samplers / chains.

The main work to be done here involves figuring out a way to handle the TemperedState such that race conditions are avoided for a multi-threaded approach, and data can be silo-ed to worker process for a distributed approach.

A multithreaded approach is probably more desirable in the short term, and its likely that a distributed approach will need more significant changes to the architecture.