Pandas 1.3.0 compatibility

[ivirshup]

We have some failing tests due to a couple bugs introduced in pandas 1.3.0:

• BUG: 1.3.0 DataFrame.agg over categorical columns with non-unique index returns wrong size result pandas-dev/pandas#42380

I think this one is small in scope. Has problems when df.agg is called, when all columns are categorical and index is non-unique. Definitely a bug in pandas, and I don't think we do this much. Switching to df.apply works around the problem.

• BUG: 1.3.0 column assignment via single columnnp.matrix behaviour change pandas-dev/pandas#42376

Assignment of single columns np.matrix to dataframe columns no longer works as if the matrix were a 1d array. I think this is a bug since it's an undocumented behaviour change. Fixes are pretty easy, since we can just wrap occurrences with np.ravel, however I wouldn't be surprised if there were many places in the codebase that this happened.

I would be good if these didn't happen.