INFER should randomly pick a *single* generator for all predictions across the whole query

[riastradh-probcomp]

Currently the user must specify the generator and let the generator produce a single summary. If the user wants to summarize the composition of two generators, that's the user's problem.

On the 20160624-riastradh-cgpm / 20160628-fsaad-cgpm branch, we pick a generator at random from the population -- but we do it differently for every prediction in a single query. Picking it at random gives the correct distribution, in aggregate, at least -- but only if we use the same generator for every INFER.

Problem: The predictions happen in independent SQL function calls across a query, so persuading them to use a common generator will take some effort.

[fsaad]

Picking it at random gives the correct distribution, in aggregate, at least -- but only if we use the same generator for every INFER -- but only if we use the same generator for every INFER.

Why?