The FlinkRunner shouldn't require all of the values for a key to fit in memory

The use of a reducer that adds all of the elements that it consumes to a list is the primary way in which this occurs - if instead, we produce a filtered iterable, or a collection of filtered iterables, we can lazily iterate over all of the contained elements without having to buffer all of the elements.

For an example of where this occurs, see `Concatenate` in  `FlinkBatchPortablePipelineTranslator`.

Imported from Jira [BEAM-4228](https://issues.apache.org/jira/browse/BEAM-4228). Original Jira may contain additional context.
Reported by: tgroh.