Calling `predict()` with a dataframe of unobserved cases does not work

[jvparidon]

Calling predict() with a dataframe containing cases with unobserved combinations of predictor values (i.e. synthetic cases) does not work.

We'll probably have to roll our own predict() for multiple membership models and make sure it handles the dummy/indicator matrix system correctly. I think this is conceptually not that difficult but working out the kinks might take a few tries.