Add `tf.contrib.bayesflow.metropolis_hastings`.

The Metropolis-Hastings accept/reject framework is useful for constructing various MCMC algorithms. Many of the MCMC algorithms are Metropolis-like, i.e., a proposal is generated and then the accept/reject procedure is performed. Current implementation accepts a user-defined target energy and proposal generating function (e.g., normal or HMC proposals) to produce a Markov Chain.

PiperOrigin-RevId: 170358662