The Java SDK Harness and runners-core-construction payloads should be consistent

Right now the Java SDK harness expects DoFnInfo instances within the payload of all DoFns it is provided, while the pipeline translation/ParDoTranslation/etc assumes that the ParDoTranslation channel will be used to extract useful values from the serialized form (potentially by converting to a ParDoLike). As such, a Java ParDo serialized with the Java ParDoTranslation class cannot be provided in that form to the Java SDK harness. Instead, that node has to be rewritten to include a `DoFnInfo` in the payload.

This seems unreasonable, and the two should be brought in line to use the same serialized representation.

Imported from Jira [BEAM-3365](https://issues.apache.org/jira/browse/BEAM-3365). Original Jira may contain additional context.
Reported by: tgroh.

Compare the Java SDK harness with runners-core-construction and the pipeline translation path, especially ParDoTranslation, DoFnInfo, and ParDoLike. Determine which serialized representation should be shared, then verify that Java ParDos can pass from translation to the harness without rewriting their payloads.

Written by the indexing model from the issue text.