Sort avoidance for group by and join

Group by and join only require that the records be grouped together by key. It is not necessary for the keys to be sorted. If we can have a Tez Input/Output implementation that does the grouping using hashmap (memory, spilling, etc have to be handled) it could really speed up group by and join. Combiners on both input and output side can also be fast if serialization/deserialization is not required and that can be used instead of POPartialAgg.