Spark runner: Combine.perKey performance issues

Combine.perKey on current implementation uses technique of creating an accumulator for each input key and then merge all these accumulators together. Optimize this by:

• changing accumulator from Iterable to Map, and using addInput as much as possible
• try to move the window explode to pre-shuffle (add window label to key for non-merging windows), measure the impact, and if the impact is substantial, implement that for at least window functions assigning to single (global) window or single window per element (tumbling windows)