[FEA] Have a repartition fallback for hash aggregates instead of sort

[revans2]

Is your feature request related to a problem? Please describe.
We know that hash partitioning is faster than sorting. In GpuHashAggregate if the intermediate values get to be too large we end up sorting the intermediate results and then doing a final pass to output the answer.

In many cases it might be good to have a heuristic in place that sees if the output got smaller so we can decide if we do #7404 instead. But in the cases when we see the output is getting a lot smaller than the input or if it is not a partial aggregate, we probably should look at doing hash partitioning of the intermediate results instead of sorting the result.

[binmahone]

is this a duplicate with #8391 ?

[revans2]

Yup I filed the same thing twice. Feel free to dupe this to the other or vise versa.

[binmahone]

Closing this as it duplicates #8391