Propagate keyed-ness of PValues properly in ReferenceRunner

ULR, just like regular direct runner, needs to respect "keyed"-ness of certain PValues, for two purposes: 1) invoking transforms of those PValues serially per key (e.g. because some of such transform evaluators, such as GABW, are not thread-safe per key), and 2) supporting state and timers.

In [https://github.com/apache/beam/pull/5585,](https://github.com/apache/beam/pull/5585,) one of the commits adds such tracking, but only naively - directly on the outputs of GBKO and GABW, without propagating downstream through key-preserving DoFn's.

This Jira is to mimic the behavior of KeyedPValueTrackingVisitor in ULR.

Imported from Jira [BEAM-4557](https://issues.apache.org/jira/browse/BEAM-4557). Original Jira may contain additional context.
Reported by: jkff.