ENH: Add full tz aware support for IntervalIndex

[jschendel]

Problem description

The fixes in #18424 allow for and IntervalIndex to be constructed with tz aware timestamps, but there are a few attributes/methods that still do not work properly with tz aware:

• Disallow an IntervalIndex to contain elements with differing tz
• IntervalIndex.mid returns tz naive
• Anything using _get_next_label or _get_previous_label will raise
  • get_indexer when passed an IntervalIndex
  • get_loc for overlapping/non-monotonic IntervalIndex
  • contains always returns False for overlapping/non-monotonic due to Try/Except
  • etc.

Code Sample, a copy-pastable example if possible

Setup:

In [2]: start = pd.Timestamp('2017-01-01', tz='US/Eastern')
   ...: index = pd.interval_range(start, periods=3)
   ...: index
   ...:
Out[2]:
IntervalIndex([(2017-01-01, 2017-01-02], (2017-01-02, 2017-01-03], (2017-01-03, 2017-01-04]]
              closed='right',
              dtype='interval[datetime64[ns, US/Eastern]]')

IntervalIndex.mid returns tz naive:

In [3]: index.mid
Out[3]:
DatetimeIndex(['2017-01-01 17:00:00', '2017-01-02 17:00:00',
               '2017-01-03 17:00:00'],
              dtype='datetime64[ns]', freq=None)

get_indexer raises when passed an IntervalIndex:

In [4]: index.get_indexer(index[:-1])
---------------------------------------------------------------------------
TypeError: cannot determine next label for type <class 'pandas.core.indexes.datetimes.DatetimeIndex'>

Differing tz:

In [5]: left = pd.date_range('2017-01-01', periods=3, tz='US/Eastern')
   ...: right = pd.date_range('2017-01-02', periods=3, tz='US/Pacific')
   ...: index = pd.IntervalIndex.from_arrays(left, right)
   ...: index
   ...:
Out[5]:
IntervalIndex([(2017-01-01, 2017-01-02], (2017-01-02, 2017-01-03], (2017-01-03, 2017-01-04]]
              closed='right',
              dtype='interval[datetime64[ns, US/Eastern]]')

In [6]: index.left.dtype
Out[6]: datetime64[ns, US/Eastern]

In [7]: index.right.dtype
Out[7]: datetime64[ns, US/Pacific]

[jschendel]

Side question: Should a plain Interval be able to hold timestamps with different tz? Seems a bit strange, but since different tz are comparable, I don't want to outright declare this a bug.

In [2]: left = pd.Timestamp('2017-01-01', tz='US/Eastern')
   ...: right = pd.Timestamp('2017-01-02', tz='US/Pacific')
   ...: iv = pd.Interval(left, right)
   ...:

In [3]: iv
Out[3]: Interval('2017-01-01', '2017-01-02', closed='right')

In [4]: iv.left
Out[4]: Timestamp('2017-01-01 00:00:00-0500', tz='US/Eastern')

In [5]: iv.right
Out[5]: Timestamp('2017-01-02 00:00:00-0800', tz='US/Pacific')

Note that an Interval cannot be constructed from an a combination of aware and naive timestamps, as the two aren't comparable:

In [6]: left = pd.Timestamp('2017-01-01', tz='US/Eastern')
   ...: right = pd.Timestamp('2017-01-01')
   ...: iv = pd.Interval(left, right)
   ...:
---------------------------------------------------------------------------
TypeError: Cannot compare tz-naive and tz-aware timestamps

[jreback]

no an interval must have he same tz in left/right (could be None in both as welll)

[jschendel]

@jreback : okay, will open a new issue for that to give it more visibility/searchability.