TFX pipelines experience a huge blowup in intermediate data size

The elements in TFX intermediate collections are dictionaries of (typically single-element) numpy arrays, which are (relatively) expensive to serialize (e.g. using pickle for the numpy wrapper of a primitive int/float, repeating the column names in every element).

Though it'd be good to use a better intermediate representation, this is exacerbated because the fusion algorithm does not pack as much possible into executable stages.

Imported from Jira [BEAM-6243](https://issues.apache.org/jira/browse/BEAM-6243). Original Jira may contain additional context.
Reported by: robertwb.