[REVIEW] Add dictionary support to libcudf groupby functions

[davidwendt]

Reference #5963 Add dictionary support to groupby.

• argmax
• argmin
• collect
• count
• max
• mean*
• median
• min
• nth element
• nunique
• quantile
• std*
• sum*
• var*

• not supported due to 10.2 compile segfault

[codecov]

Codecov Report

Merging #6585 (6cc32d8) into branch-0.18 (31c0d29) will increase coverage by 0.02%.
The diff coverage is n/a.

@@               Coverage Diff               @@
##           branch-0.18    #6585      +/-   ##
===============================================
+ Coverage        82.09%   82.11%   +0.02%     
===============================================
  Files               97       97              
  Lines            16474    16477       +3     
===============================================
+ Hits             13524    13530       +6     
+ Misses            2950     2947       -3     

Impacted Files	Coverage Δ
python/cudf/cudf/_fuzz_testing/fuzzer.py	0.00% <0.00%> (ø)
python/cudf/cudf/utils/hash_vocab_utils.py	100.00% <0.00%> (ø)
python/cudf/cudf/core/abc.py	91.48% <0.00%> (+4.25%)	⬆️
python/cudf/cudf/utils/gpu_utils.py	58.53% <0.00%> (+4.87%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 31c0d29...6cc32d8. Read the comment docs.

[jrhemstad]

I realized that this PR should wait until #6392 as it will likely have a lot of conflicts.

[davidwendt]

I cannot get around the 10.2 ptxas compile segfault as documented here https://nvbugswb.nvidia.com/NvBugs5/SWBug.aspx?bugid=3186317
The code logic that seems to cause the error is isolated in the aggregation sum specialization for dictionary type. This means we cannot support aggregation-sum-groupby for dictionary columns as values. And aggregation-sum is also required in order to support mean, std, and var aggregations. Using a dictionary column as keys for groupby work fine.

The PR includes logic to throw an exception if a dictionary column is used as values with one of these aggregation types. The code is isolated around a CUDA==10.2 equivalent compile directive so technically these aggregations work if compiled with 10.1 or 11.0 but the gtests have been commented out.

[davidwendt]

rerun tests

[mike-wendt]

rerun tests

[devavret]

Damn. I missed it by a minute.

[devavret]

This is sort of the reverse direction of what I've been doing. column_device_view's element() accessor is more helpful than simple subscript. e.g. in fixed point. element() will return the value with the scale applied. And that scale is stored once in the column.

Any particular reason for this change?