BUG: Bug in .loc performing fallback integer indexing with object dtype indices (GH7496)

[jreback]

closes #7496

In [4]: s = Series()

In [5]: s.loc[1] = 1

In [6]: s.loc['a'] = 2

In [7]: s.loc[-1]
KeyError: 'the label [-1] is not in the [index]'

In [8]: s.loc[[-1, -2]]
Out[8]: 
-1   NaN
-2   NaN
dtype: float64

In [9]: s.loc[['4']]
Out[9]: 
4   NaN
dtype: float64

In [10]: s.loc[-1] = 3

In [11]: s.loc[[-1,-2]]
Out[11]: 
-1     3
-2   NaN
dtype: float64

In [12]: s['a'] = 2

In [13]: s.loc[[-2]]
Out[13]: 
-2   NaN
dtype: float64

In [14]: del s['a']

In [15]: s.loc[[-2]] = 0
KeyError: '[-2] not in index'

[code-of-kpp]

s.loc[[-2]] = 0

shouldn't -2 to be added to index after that?

[code-of-kpp]

this one looks similar #5420

[jreback]

for the s.loc[[-2]] = 0.

Setting with a single indexer (e.g. s.loc[-2] = 0 is fine as this is enlargement.
However, setting with a list (either on a Series or a DataFrame) was deemed to be
more 'error prone', so it was disallowed.

e.g. both s.loc[[-2]] = 0 and df.loc[[-2]] = 0 are not allowed unless the location already exists.
though s.loc[-2] = 0 and df.loc[-2] = 0 will work