Numpy versions < 1.8 play poorly with arrays of mixed type, truncating anything not of type string to the length of the longest string. This is inhibiting #7974

So the issue is as follows

>>> import numpy as np
>>> np.__version__
'1.7.1'
>>> np.array(['A', 'A', np.nan, 'B', -np.inf, 3.14], dtype=np.unicode)
array([u'A', u'A', u'n', u'B', u'-', u'3'], 
      dtype='<U1')

two possible solutions:

1. Fold back in shims removed in Remove numpy 1.6 specific work-arounds  #7484 that existed mainly to solve this problem
2. throw a warning if category.py is imported with a numpy<1.8 and skip tests for numpy<1.8

The advantage of 2 is it keeps the code cleaner and this is an edge case of sorts because it's only an issue when the array is of mixed type (strings and numbers). The advantage of 1 is of course edge case support on older systems.