BUG: nanops._has_infs doesn't work with many dtypes

[toddrjen]

nanops._has_infs, which is used in determining whether to employ bottleneck optimizations, fails on dtypes which should otherwise be supported, including 'f2', complex, and multidimensional 'f4' and 'f8'.

>>> from pandas.core import nanops
>>> import numpy as np
>>> 
>>> val0 = np.ones(10)*np.inf
>>> val1 = val0.astype('f2')
>>> val2 = val0+val0*1j
>>> val3 = np.tile(val0, (1, 10))
>>> 
>>> nanops._has_infs(val0)
True
>>> nanops._has_infs(val1)
False
>>> nanops._has_infs(val2)
False
>>> nanops._has_infs(val3)
ValueError: Buffer has wrong number of dimensions (expected 1, got 2)

[jreback]

FYI, dtype f2 is not ALWAYS available on all platforms/numpies (nor are some complex numbers)

you can test if the dtype.name is in _TYPE_MAP (if its numeric), see here: https://github.com/pydata/pandas/blob/master/pandas/src/inference.pyx

[jreback]

IIRC bottleneck does NOT support f2 (not sure though)

[toddrjen]

The current implementation has 2 major problems:

1. It assumes the array is 1D
2. It assume any array that is not an f4 or f8 dtype cannot have infinite values

These problems are both easily fixed without much additional performance overhead.

[jreback]

closed by #7448