DataFrame constructor is inconsistent when coercing values to strings with `dtype=str`.

[TomAugspurger]

When providing dtype=str to the DataFrame constructor, we're inconsistent about coercing values to strings.

When there's no overlap between the keys of data and columns, things are probably OK.

In [4]: pd.DataFrame(index=[0, 1], columns=[0, 1], dtype=str)
Out[4]:
     0    1
0  NaN  NaN
1  NaN  NaN

(those values are np.nan).

But when there is an overlap between keys of data and columns, the newly introduced values are coerced to strings.

In [8]: pd.DataFrame({'A': [1, 2]}, index=[0, 1], columns=['A', 'B'], dtype=str)
Out[8]:
   A    B
0  1  nan
1  2  nan

(everything in that dataframe is a string, like "1" or "nan")

That's be cause init_dict relies on arrays_to_mgr to coerce the values to the dtype, and arrays_to_mgr only gets a single dtype.

[TomAugspurger]

This is not fixed in #24387

[FR4NKESTI3N]

I am not able to reproduce this in pandas 0.23.4:

>>> pd.DataFrame({'A': [1, 2]}, index=[0, 1], columns=['A', 'B'], dtype=str)['B'][0] is np.nan
True

[TomAugspurger]

Huh I must have been on the wrong branch again. Thanks @FR4NKESTI3N.