In linear_pool_sample, standardize the compound unit or "joint across variables"

[elray1]

Suppose that model A gets samples from a joint distribution across horizon (trajectories), and model B gets samples from a joint distribution across both horizon and location. Current ensemble functionality for samples would end up with a mix of samples that handle different types of dependency. It would be good to standardize to the "lowest common denominator" so that all ensemble samples have output_type_id values encoding the same type of dependence structure.