BUG: Index.union cannot handle array-likes

[sinhrks]

Closes #10149.

• Added explicit tests for array-likes in set-ops, intersection, union, difference and sym_diff.
  • TimedeltaIndex previously did additional input check which is inconsistent with others. Made it work as the same manner as others.
  • MultiIndex set-ops should only accept list-likes of tuples. Implement the logic and add explicit test.
• Changed to understandable error message for str which has __iter__ in py3.

# current error msg
idx.intersection('aaa')
# TypeError: Index(...) must be called with a collection of some kind, 'aaa' was passed

NOTE: This DOESN'T care the name attribute checks, which is being worked in #9965.

[jreback]

can you move these to Base (so that they get tested for all types)?

[jreback]

rebase / address comments when you have a chance

[sinhrks]

Rebased. I've added tests to Base, but it cannot work in all Index currently. I'll prepare separate issue about CategoricalIndex and TimedeltaIndex set ops behavior.

What I'd like to confirm here is MultiIndex should accept list-likes of something? Currently, it raises error not understandable for users as below.

>>> midx = pd.MultiIndex.from_tuples([(1, 2), (3, 4)])
>>> midx
MultiIndex(levels=[[1, 3], [2, 4]],
           labels=[[0, 1], [0, 1]])
>>> midx.intersection([(1, 2)])
AttributeError: 'list' object has no attribute 'names'

[sinhrks]

Let me update TimedeltaIndex behavior here (because simple than I initially thought). On current master, DatetimeIndex.union performs additional check and it causes an error. Removed this check in and made it consistent with others in this PR.

idx1 = pd.timedelta_range('1 day', freq='D', periods=5)
idx2 = pd.timedelta_range('3 day', freq='D', periods=5)

idx1.intersection(idx2.asobject.tolist())
# TimedeltaIndex(['3 days', '4 days', '5 days'], dtype='timedelta64[ns]', freq=None)

idx1.union(idx2.asobject.tolist())
# AttributeError: 'list' object has no attribute 'inferred_type'

idx1.difference(idx2.asobject.tolist())
# TimedeltaIndex(['1 days', '2 days'], dtype='timedelta64[ns]', freq=None)

idx1.sym_diff(idx2.asobject.tolist())
# TimedeltaIndex(['1 days', '2 days', '6 days', '7 days'], dtype='timedelta64[ns]', freq=None)

[sinhrks]

Rebased, and updated top descriptions about TimedeltaIndex and MultiIndex behaviors.

[jreback]

thanks @sinhrks