Remove null-splitting from _perform_aggregation

[charlesbluca]

Duplicate of #273, which was merged in before some groupby split_out tests could be resolved

[charlesbluca]

Now that we are able to access dask-cudf's groupby agg codepath, the task graph size varies depending on if we use GPU or not. There could be a clean way to parametrize this all into one test, but I opted for this for now to make the difference more clear.

In general, is there a way we can verify that split_every is working that doesn't depend on checking the task graph? Asking because now upstream changes in dask-cudf can break this test.

[codecov-commenter]

Codecov Report

Merging #326 (393f11c) into main (a3fc92d) will decrease coverage by 0.06%.
The diff coverage is 86.66%.

@@            Coverage Diff             @@
##             main     #326      +/-   ##
==========================================
- Coverage   95.80%   95.74%   -0.07%     
==========================================
  Files          65       65              
  Lines        2834     2841       +7     
  Branches      426      429       +3     
==========================================
+ Hits         2715     2720       +5     
- Misses         74       75       +1     
- Partials       45       46       +1     

Impacted Files	Coverage Δ
dask_sql/physical/rel/logical/aggregate.py	94.80% <85.71%> (-1.12%)	⬇️
dask_sql/physical/utils/groupby.py	100.00% <100.00%> (ø)

---

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 a3fc92d...393f11c. Read the comment docs.