float32 has priority over float64 for scalar integer ufunc inputs with other arrays

[pv]

Consider this:

>>> from scipy.special import sph_harm
>>> sph_harm.types
['llff->F', 'lldd->D', 'ffff->F', 'dddd->D']
>>> sph_harm(0,[0],0,0).dtype
dtype('complex64')   # <- uh-oh
>>> sph_harm(0,[0],0,0.0).dtype
dtype('complex128')
>>> sph_harm(0,0,0,0).dtype
dtype('complex128')
>>> sph_harm(0,[0],[0],0).dtype
dtype('complex128')

Apparently, the ufunc loop selector prefers to cast integers to float32 in this case, which is somewhat surprising. This may have to do with the fact that given an integer array as one of the first two arguments, there are still two possible loop choices, and deciding which to use would need considering appropriate scalar casting rules for the last two arguments. Note that if one of the last arguments is an array instead of a scalar, the dtype decision is as expected, so maybe this is an issue with scalar special cases.