Implement Combine optimizations for GABW in streaming

This should be straight-forward.
Introduce `AccumT` generics in `SparkGroupAlsoByWindowViaWindowSet` and call with `InputT` for GBK and `AccumT` with Combine.
Pass the proper `SystemReduceFn` instead of creating it in `SparkGroupAlsoByWindowViaWindowSet`.
For combine, extract the output from the fired accumulated output.

Imported from Jira [BEAM-1591](https://issues.apache.org/jira/browse/BEAM-1591). Original Jira may contain additional context.
Reported by: amitsela.

Start by locating SparkGroupAlsoByWindowViaWindowSet and tracing how GABW handles GBK and Combine inputs. Review the existing SystemReduceFn construction and fired accumulated output flow. Done means AccumT is used for Combine, the appropriate SystemReduceFn is passed in, and Combine output is extracted from the fired accumulated output.

Written by the indexing model from the issue text.