BUG: Arrow duration reduction consistency

[rhshadrach]

When using .sum() on duration[ns][pyarrow] we get a Timedelta, but all other units give datetime.timedelta.

import datetime as dt

ser = pd.Series([dt.timedelta(seconds=1)], dtype="duration[ns][pyarrow]")
print(type(ser.iloc[0]), type(ser.sum()))
# <class 'pandas.Timedelta'> <class 'pandas.Timedelta'>

ser = pd.Series([dt.timedelta(seconds=1)], dtype="duration[us][pyarrow]")
print(type(ser.iloc[0]), type(ser.sum()))
# <class 'pandas.Timedelta'> <class 'datetime.timedelta'>

ser = pd.Series([dt.timedelta(seconds=1)], dtype="duration[ms][pyarrow]")
print(type(ser.iloc[0]), type(ser.sum()))
# <class 'pandas.Timedelta'> <class 'datetime.timedelta'>

ser = pd.Series([dt.timedelta(seconds=1)], dtype="duration[s][pyarrow]")
print(type(ser.iloc[0]), type(ser.sum()))
# <class 'pandas.Timedelta'> <class 'datetime.timedelta'>

I suspect we want to make all of the sums be Timedeltas. I haven't investigated other reductions, but suspect we'll find similar behavior.

cc @jbrockmendel

[ahalsey]

This needs some additional testing, so I'm leaving my fix in a draft state for the moment and plan to circle back later today.

[mroeschke]

This may be a pyarrow specific behavior, xref apache/arrow#34614, but we can wrap in a pandas.Timedelta for consistency

[jbrockmendel]

Yah id expect these all to give pd.Timedelta