fix grouped nanmean with bool

[dcherian]

numbagg.grouped.group_nanmean(np.array([True, True, False]), [0, 0, 0], axis=-1)
# array([0], dtype=int32)
np.nanmean(np.array([True, True, False]))
# 0.6666666666666666

Should I just promote these to float instead?

[dcherian]

Another edge case

numbagg.grouped.group_nansum(np.array([True, True, False]), [0, 0, 0], axis=-1).dtype
# int32
np.nansum(np.array([True, True, False]), axis=-1).dtype
# int64

[max-sixty]

This is somewhat downstream of #121. There's a bullet tucked away there "Always keep types — current state (except for bools...)"

The current state is that we maintain types, apart from a quirk where we convert bools to int32.

I think the closest thing would be for flox to convert into whatever type it wants before calling numbagg, and then these aren't dependent on any handling in numbagg. Then later we can optimize by solving #121. WDYT? Or is this more complicated than converting bools to floats and flox needs to then handle lots of edge cases?

[dcherian]

Yeah, I'll just promote for nanmean now. I think that's easiest.