BUG: DatetimeArray/Series incorrectly accepts timedelta64('NaT') for __setitem__

[jbrockmendel]

dti = pd.date_range('2016-01-01', periods=3)
nat = np.timedelta64('NaT')

dta = dti._data.copy()
dta[0] = nat  # <-- should raise TypeError, doesn't

ser = pd.Series(dti.copy())
ser[1] = nat  # <-- should raise TypeError, doesn't

df = pd.DataFrame(dti.copy())
df.loc[2, 0] = nat  # <-- should raise TypeError, doesn't

I haven't checked, but I expect the same thing happens in reverse for Timedelta, possible Period dtypes.

[jbrockmendel]

"should" behavior in OP is wrong, it should cast to object dtype, but does does not

[jbrockmendel]

I think I incorrectly marked #27311 as closing this, while that only fixes the Series portion of this. I still need to fix the DTA version.

[jbrockmendel]

Woops, #27331 does fix the other half of this.