(Datetime|Timedelta)Index +/- np.nan

[jbrockmendel]

AFAICT this is now the last thing standing between us and getting rid of ops._Op and ops._TimeOp. Series ops currently case np.nan to pd.NaT whereas Index ops treat it as a float and raise a TypeError.

tdi = pd.TimedeltaIndex(['1 Day'])
ser = pd.Series(tdi)

>>> tdi + np.nan
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: unsupported operand type(s) for +: 'TimedeltaIndex' and 'float'

>>> ser + np.nan
0   NaT
dtype: timedelta64[ns]

Note also that the behavior of Series[datetime64] has changed since 0.22.0:

0.22.0:

dti = pd.date_range('2016-01-01', periods=2)
ser = pd.Series(dti)

>>> dti - np.nan
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: unsupported operand type(s) for -: 'DatetimeIndex' and 'float'

>>> ser - np.nan
0   NaT
1   NaT
dtype: timedelta64[ns]

# Both raise on addition, assuming that np.nan gets cast to pd.NaT wearing its datetime hat.

ATM ser - np.nan also raises, since that now delegates to the DatetimeIndex op.