Implementing variance reduction

[mikkelbue]

Variance reduction (VR) for MLDA has not been implemented yet. Please refer to the paper for details.

Implementing this would require:

• Allowing for randomising the length of the subchain used to produce a proposal to the next-higher level.
• (For example) writing a function that would extract the chains from every level to an arviz.InferenceData object and then apply VR to the appropriate links across levels. To find the appropriate links, we would probably have to cache the history of which links were promoted to proposals.