2.1: Structure of generic MCMC samplers

Created by: gcarleo

We should define a generic TransitionKernel - based sampling for MCMC, see for example this nice implementation in Tensorflow. At the same time, we can remove all the duplicate parallel tempering samplers, as done in tfp.mcmc.ReplicaExchangeMC.

Notice that tfp.mcmc.sample_chain returns the states in the generated chain but doesn't store (as far as I can tell) neither the log of the sampled probabilities nor its derivatives