prob / se prediction with ensemble pipeops

[mb706]

There should probably be options how the prob and se columns are calculated. Currently we just average them, but there are probably mutliple settings in which ensembling happens, and the requirements for how se / prob should be calculated might differ.

An example is multiple imputation. In that case, the se prediction should be the sqrt of (mean of se^2 + some factor times variance of means).