DataFrame API: groupby(dropna=False) still drops NAs when grouping on multiple columns or indexes

df.groupby(['foo', 'bar'], dropna=False).sum()

This will still drop NAs in the output.

This is due to pandas bug 36470 "BUG: groupby(..., dropna=False) excludes NA values when grouping on MultiIndex levels".

We implement groupby by moving all grouped data into the index and requiring Index() partitioning, so we will always run into this issue, even when the user is grouping on columns, not indexes.