BUG: nanops._ensure_numeric can't handle object dtypes containing complex objects

[toddrjen]

nanops._ensure_numeric is used to make sure ndarrays are a numeric dtype. However, it can't handle the case where the dtype is object and one or more of the axes is complex:

>>> from pandas.core.nanops import nanmean
>>> import numpy as np
>>> 
>>> value = np.vstack([np.array(np.nan).astype('O'), np.array(1+1j).astype('O')])
>>> nanmean(value, axis=0)
TypeError: can't convert complex to float

[jreback]

I marked this a bug, but this violates the guarantees of the nan functions. The dtypes on a single axis need to be in the proper converted dtype already or passed as a single dtype block (multi-dim). Since this is an internal function it shouldn't be responsible for conversions. object dtypes are by definition non-convertible (internally).

The ensure_numeric is really not necessary and is old code for compat (when things weren't converted), or as a convience.

In [7]: DataFrame(np.vstack([np.array(np.nan).astype('O'), np.array(1+1j).astype('O')])).dtypes
Out[7]: 
0    object
dtype: object

In [11]: DataFrame(np.vstack([np.array(np.nan).astype('O'), np.array(1+1j).astype('O')])).mean()
Out[11]: 
0    (1+1j)
dtype: complex128

In [9]: DataFrame(np.vstack([np.array(np.nan), np.array(1+1j)])).dtypes
Out[9]: 
0    complex128
dtype: object

In [10]: DataFrame(np.vstack([np.array(np.nan), np.array(1+1j)])).mean()
Out[10]: 
0    (1+1j)
dtype: complex128

In [12]: DataFrame(np.vstack([np.array(np.nan).astype('O'), np.array(1+1j).astype('O')])).dtypes
Out[12]: 
0    object
dtype: object

[toddrjen]

This is fixed as a side-effect of #7448