common._isnull_ndarraylike(...) uses lib.isnullobj to check
nulls/NaN/NaT in ndarray, which in turn relies on util._checknull.
_checknull did not know about NaT, but now lib.isnullobj does, while
still maintaining performance by doing arr[i] only once.

Added a test case test_isnull_nat() to test_common.py and
check for NaT in lib.isnullobj. pd.isnull(np.array([pd.NaT]))
now yields the correct results ([True]).