BUG: Only convert 0d arrays to scalars in power - #4146

It might be possible to simplify the function using other conversion functions, but this fixes the issue at hand nicely.

This doesn't fix the problem:

$ python -W always
Python 2.7.6 |Anaconda 1.8.0 (64-bit)| (default, Nov 11 2013, 10:47:18) 
[GCC 4.1.2 20080704 (Red Hat 4.1.2-54)] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import numpy as np
>>> np.__version__
'1.9.0.dev-7194978'
>>> x = np.array([1, 2, 3])
>>> y = np.array([[2]])
>>> x ** y
__main__:1: DeprecationWarning: converting an array with ndim > 0 to an index will result in an error in the future
array([1, 4, 9])

If I don't use -W always, I don't get the DeprecationWarning (as expected), but the shape of the result is still wrong.

This PR only works if I run python as python -W error (or use warnings.filterwarnings to do the same thing).

$ python -W error
Python 2.7.6 |Anaconda 1.8.0 (64-bit)| (default, Nov 11 2013, 10:47:18) 
[GCC 4.1.2 20080704 (Red Hat 4.1.2-54)] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import numpy as np
>>> x = np.array([1, 2, 3])
>>> y = np.array([[2]])
>>> x ** y
array([[1, 4, 9]])
>>> 

OOps, right point, completly nonsense fix :)

Updated, could remove the numpy scalar branch, since we support Index, but it should not really matter either way.
EDIT: nvm about that, that would require some other handling for the float scalars.

Maybe a bit late for 1.8.1rc1, but this would be useful to merge and backport. Just ran into this, and it's quite annoying to avoid.

OK, lets give it a shot. I supose 1.8.1 will give us some testing of these fixes also.

Merged, thanks Sebastian.

Oops, needs a rebase now.

Rebased.

Why does the test still need to filter warnings? The code being tested should not raise a warning. If it does, it is a bug.

Oops, yes, it does not do that anymore, so removed it.

merged but its a nontrivial backport, if you want it in do it quick