Fix Merge divisions after filtering partitions

[rjzamora]

A Dask-cuDF user reported an error that suggested something was going wrong in the Merge graph:

kwargs:    {'how': 'inner', 'indicator': False, 'left_index': False, 'right_index': False, 'suffixes': ('_x', '_y'), 'result_meta': Empty DataFrame
Columns: [id_x, id_y, jaccard, id, uid]
Index: [], 'left_on': ['id_y'], 'right_on': ['id']}
Exception: 'AttributeError("\'tuple\' object has no attribute \'merge\'")'

It turns out that the problem happens when you perform a broadcast merge after filtering the partitions of the larger collection. In main, we use _divisions() to check the divisions of a (possibly-filtered) child expression when I'm pretty sure we need to be using the child's divisions property. This PR makes the necessary fix and adds test coverage.

[phofl]

can you check the result as well?

[rjzamora]

Good call.

While adding this check I actually noticed that divisions are often "wrong" after a merge (unrelated to the changes in this PR). More specifically, it seems like we are inheriting divisions after a broadcast join, even if we aren't merging on the index. If I understand correctly, we inherit the partition count in this case, but the divisions are likely to change.

[phofl]

Yeah that's correct, we probably should only inherit on index merges, pandas is not consistent with the index when you merge on columns :(

[phofl]

thanks