CI: Fix tests broken by np 1.18 sorting change

[jbrockmendel]

Longer-term, do we want to use numpy's new behavior?

cc @TomAugspurger

Update A little more background: consider two ndarrays in np<1.18

arr1 = np.array([1, 2, np.nan, 3, 4])
arr2 = np.array([1, 2, np.datetime64("NaT"), 3, 4], dtype="datetime64[ns]")

>>> np.sort(arr1)
array([ 1.,  2.,  3.,  4., nan])
>>> np.sort(arr2)
array([                          'NaT', '1970-01-01T00:00:00.000000001',
       '1970-01-01T00:00:00.000000002', '1970-01-01T00:00:00.000000003',
       '1970-01-01T00:00:00.000000004'], dtype='datetime64[ns]')

In numpy 1.18, the behavior of np.sort(arr2) is changing to put the NaT at the end instead of at the beginning, to behave more like NaN. This breaks a few tests, which this PR fixes.

Side-note: as of now, 1.18 changes the behavior for datetime64, but not timedelta64.

[jorisvandenbossche]

Longer-term, do we want to use numpy's new behavior?

We already do? (put NaT last)

In [5]: pd.Series(np.array([1, 2, np.datetime64("NaT"), 3, 4], dtype="datetime64[ns]")).sort_values()  
Out[5]: 
0   1970-01-01 00:00:00.000000001
1   1970-01-01 00:00:00.000000002
3   1970-01-01 00:00:00.000000003
4   1970-01-01 00:00:00.000000004
2                             NaT
dtype: datetime64[ns]

In [6]: np.sort(np.array([1, 2, np.datetime64("NaT"), 3, 4], dtype="datetime64[ns]")) 
Out[6]: 
array([                          'NaT', '1970-01-01T00:00:00.000000001',
       '1970-01-01T00:00:00.000000002', '1970-01-01T00:00:00.000000003',
       '1970-01-01T00:00:00.000000004'], dtype='datetime64[ns]')

[jorisvandenbossche]

We already do? (put NaT last)

Ah, I tested for Series. But for Index we indeed follow numpy's behaviour and put it first.

[jorisvandenbossche]

Going to merge this to get CI green again.

Longer-term, do we want to use numpy's new behavior?

@jbrockmendel can you open an issue for this?

Side-note: as of now, 1.18 changes the behavior for datetime64, but not timedelta64.

Known, or something to report to numpy?