BUG: IntervalIndex([np.nan, np.nan]).is_monotonic returns True

[jbrockmendel]

I'm pretty sure this is related to other questionable behavior:

index = pd.IntervalIndex([np.nan, np.nan])
other = pd.IntervalIndex([np.nan])

In [4]: index.is_monotonic  # should be False
Out[4]: True

In [5]: index._index_as_unique  # should be False
Out[5]: True

In [6]: index.get_indexer_for(other)  # shouldn't raise
InvalidIndexError: Reindexing only valid with uniquely valued Index objects

[jbrockmendel]

@jschendel I've got a branch trying to fix this and could use your input as ive broken a bunch of test_is_overlapping cases.

The main thing ive done is in IntervalTree.__init__ defined self._na_count = len(mask) - mask.sum(), then 1) edited is_monotonic to return False if self._na_count > 0 and 2) edited is_overlapping to return True if self._na_count > 1

Finally I've edited IntervalIndex._get_indexer_pointwise to handle np.nan

         indexer, missing = [], []
         for i, key in enumerate(target):
+            if is_interval_dtype(target.dtype) and isna(key):
+                # self.get_loc(np.nan) will treat it as a float instead of as
+                #  our own dtype.
+                # TODO: handle this in get_loc?
+                locs = self.isna().nonzero()[0]
+                if len(locs) == 0:
+                    missing.append(i)
            else:
                  [what we have now, indented one more time]

This suffices to fix the three things in the OP, but breaks is_overlapping tests, so I'd like to get your thoughts on if this is the right way to solve this. (also how to keep get_indexer consistent with get_loc)