Support zero-shuffle grouping operations

On some occasions input dataset might be already correctly shuffled (i.e. as a result of previous operation(s)), which means that subsequent grouping operation could leverage this and remove the unneeded shuffle. Example (pseudocode):

 Dataset<Integer> input = ...

 Dataset<Pair<Integer, Long>> counts1 = CountByKey.of(input)

   .keyBy(e -> e)

   .windowBy( /* some small window */ )

   .output();

 Dataset<Pair<Integer, Long>> counts2 = SumByKey.of(counts1)

   .keyBy(Pair::getFirst)

   .windowBy( /* larger window */ )

   .output();

Now, the second ReduceByKey already might have correct shuffle (depends on runner), but isn't able to leverage this, because it isn't aware that the key grouping key has not changed from the previous operation.

Proposed change:

 Dataset<Integer> input = ...

 Dataset<Pair<Integer, Long>> counts1 = CountByKey.of(input)

   .keyBy(e -> e)

   .windowBy( /* some small window */ )

   .output();

 Dataset<Pair<Integer, Long>> counts2 = SumByKey.of(counts1)

   .keyByLocally(Pair::getFirst)

   .windowBy( /* larger window */ )

   .output();

Introduce keyByLocally to keyed operations, which will tell the runner that the grouping is preserved from one keyed operator to the other.

This will probably require some support on Beam SDK side, because this information has to be passed to the runner (so that i.e. FlinkRunner can make use of something like DataStreamUtils#reinterpretAsKeyedStream.