[FEA] Hash Aggregate Cleanup to make closer to spark

[revans2]

Is your feature request related to a problem? Please describe.
A lot of the processing in the GPU aggregate operations are not as close to spark as they could be. It would be nice to make it more similar to help reduce the possibility of bugs and to increase the possibility of code reuse.

[abellina]

It seems like this issue can be closed by #3910.

We have made ourselves closer (re: setupReferences), and our interface calls out the relationship between our work and aggBufferAttributes in Spark. I am not sure how much closer we could get to the CPU. But adding the comment to this issue to see if we are satisfied with #3910 or we need more work to get us closer.

[abellina]

This issue should have been closed by the work in #3910 and #4272.

In a nutshell, the aggregate.scala code was made simpler and more consistent between group-by and reduction, in addition to removing complicate code in setupReferences that was not required. In AggregateFunctions.scala the GpuAggregateFunction and CudfAggregate interfaces were better defined and documented and flexible enough for complicated operations like overflow checking for decimals.