latent variables cannot be inputs to CGPM foreign models

create generator g for p using cgpm (
    model foo given err using venturescript (sp = errfoo),
    latent err numerical
)

This will not work because cgpm_metamodel._create_schema assumes that every input to a foreign model exists as a manifest variable of the population, and passes them to core.bayesdb_variable_stattype to determine the statistical type.

The logic in _create_schema must be changed to defer whatever processing requires the statistical types of inputs to foreign models until after all clauses have been ingested, and then update that processing to use latents (the mapping from latent variable names to their respective statistical types) instead of core.bayesdb_variable_stattype when appropriate.