Support for "non-trivial" ensembles of samples

[elray1]

What do we want to do about ensembles of samples where the contributing models are not weighted equally, or where they don't provide equal numbers of samples? One proposal was outlined under "Case 2" in this comment. However, in a discussion in a devteam meeting we had some questions about whether those methods would be what we would really recommend for use by people building ensembles.

Let's use this discussion as a place to settle on the methods we would actually recommend.