[REVIEW] Add COLLECT groupby aggregation

[shwina]

Closes #5620

• Adds groupby COLLECT agg to libcudf
• Python bindings

[GPUtester]

Please update the changelog in order to start CI tests.

View the gpuCI docs here.

[harrism]

@shwina moving to 0.16 since still WIP.

[rgsl888prabhu]

C++ LGTM

[jrhemstad]

Is this actually needed? The target_type logic is for determining the type to use for an accumulator for ops like sum/min/max. I wouldn't think it would be needed for COLLECT.

[shwina]

Yes -- otherwise we fail this check: https://github.com/rapidsai/cudf/blob/branch-0.16/cpp/src/groupby/groupby.cu#L102

which eventually calls:

cudf/cpp/include/cudf/detail/aggregation/aggregation.hpp

Line 528 in 5fbffb6

return (not std::is_void<target_type_t<Source, k>>::value);

[mythrocks]

Sorry for not having looked at this PR sooner. Thanks for working on this. This is something the Spark team would be interested in as well. :]

@shwina, perhaps at a later date, we might consider a test for collecting struct columns. We can try adding one when we attempt collect as a rolling-window aggregation.

[shwina]

rerun tests