Refactor all "metamodels" to return a list of answers, one per model, rather than aggregation.

[fsaad]

All aggregation over cgpms in a metamodel should be implemented by the BQL interpreter. The interface for IBayesDBMetmaodel should thus return a list of answers, rather than a single answer. This interface allows us to

• Implement ESTIMATE ... FROM <pop> with only minor modification, migrating the aggregation logic over a models of a generator into bqlfn.py (in a similar way that bqlfn.py current aggregates over metamodels in a population).
• Implement SIMULATE <bql> FROM MODELS OF <pop> which we cannot do currently because the USING MODEL support has been removed. Even if we had USING MODEL we would be obliged to run the queries row-by-row in the SQL style, rather than use multiprocessing over gpms.

Tasks:

• [ ] Update methods of cgpm_metamodel to pass through the list of results from engine to their caller.
• [ ] Update methods of crosscat to pass through a singleton list of results, since the interface of lovecat does not
• [ ] Update all invokers of metamodel.method (found using grep, typically in bqlfn) to aggregate appropriately over the returned lists, rather than assume a float has been returned.

[fsaad]

The bulk of the work for this ticket has been implemented in 4d6e0ce176538f5c5c939e8fbcd5438d116a0667

We still need to implement these stubs https://github.com/probcomp/bayeslite/blob/master/src/bql.py#L845-L851 although those are now separate smaller tickets.