BUG: ndarray.__pow__ does not check result of fast_scalar_power

[mhvk]

Backport of #9112.

While testing astropy with array_ufunc, I found the puzzling behaviour that logarithmic quantities, which cannot be raised usefully to any power, gave SystemError when raising to 0.5. After looking into this, the reason is that if one raises an ndarray or subclass to particular powers, fast_scalar_power is used in number.c -- this defers to another function (sqrt in this case), but does not check whether an error occurred. Now if a subclass raises an error, and returns NULL, that gets interpreted as meaning the fast operation is not possible, and power is tried. E.g.,

class A(np.ndarray):
def array_ufunc(self, ufunc, method, *inputs, **kwargs):
print(ufunc)
raise ValueError

a = np.ones(1).view(A)
a ** 0.5

yields

<ufunc 'sqrt'>
<ufunc 'power'>---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
in array_ufunc(self, ufunc, method, *inputs, **kwargs)
3 print(ufunc)
----> 4 raise ValueError
5

ValueError:

During handling of the above exception, another exception occurred:

SystemError Traceback (most recent call last)
in ()
6
7 a = np.ones(1).view(A)
----> 8 a ** 0.5

in array_ufunc(self, ufunc, method, *inputs, **kwargs)
1 class A(np.ndarray):
2 def array_ufunc(self, ufunc, method, *inputs, **kwargs):
----> 3 print(ufunc)
4 raise ValueError
5

SystemError: <built-in method write of _io.TextIOWrapper object at 0x7f1307c7b630> returned a result with an error set

[charris]

Thanks Marten.