Fix upstream

[jsignell]

• Closes Upstream failures with idxmaxmin #7535
• Tests added / passed
• Passes black dask / flake8 dask / isort dask

Something might have changed in pandas groupby code, but this seems like a good solution for cases where there might be only one partition in a tree reduction. I didn't check to see if there are more cases where this pattern needs updating.

[jsignell]

Good with you @jrbourbeau? I grepped around for groupby(level=0) and reset_index(level=1 and I don't see anywhere else that would have the same issue. So I think this can be merged and doesn't need to be treated as an indicator of a larger issue.

[jrbourbeau]

Sorry for not getting to getting to this sooner. I think the len(x) == 0 case was to handle empty DataFrames. Do you know if our test suite covers the len(x) == 1 case? I'm somewhat surprised this works if we just pass a non-empty DataFrame through

[jsignell]

I think in the tests that were failing, the issue was that in the agg step, len(x) was occasionally == 1. This happened when split_every=2 and the total number of partitions was 3, so sometimes there was only one item in x.