PERF: Groupby.idxmax

[jbrockmendel]

cc @phofl this is currently failing locally and could use a fresh pair of eyes.

[phofl]

I'll try to take a look over the weekend

[jbrockmendel]

@phofl im mothballing this to clear the queue, just a ping so this doesn't fall off your radar entirely

[phofl]

I think the main problem is that _wrap_aggregated_output isn't working for unordered categories that are empty. All other reductions that are tested are raising TypeErrors, while idxmax and idxmin return the wrong result. Any ideas what we could do there?

[rhshadrach]

There is a bit of an oddity with how missing observations are handled. When there are multiple groupings, we do not include the unobserved categories in e.g. grouper.result_index and fill in any unobserved ones in _wrap_aggregated_output. However, if there is just a single grouping we do include the unobserved categories in e.g. grouper.result_index and so we don't fill them in later on. This makes this approach not work for some cases of categoricals.

I plan to look into making it so we never include the unobserved categories until _wrap_aggreagted_output in the single grouping case. If we can make that work, this approach would work.