Scalar compared with an obj implementing __array_ufunc__ becomes 0d-array

Behaviors for comparison operators and for arithmetic operators are different.

Reproducing code example:

import numpy as np
class C(object):
    def __array_ufunc__(self, *args, **kwargs): print(args); print(kwargs)

np.float32(1) + C()
np.float32(1) < C()

output:

(<ufunc 'add'>, '__call__', 1.0, <__main__.C object at 0x114d54048>)
{}
(<ufunc 'less'>, '__call__', array(1., dtype=float32), <__main__.C object at 0x1145d5d30>)
{}

Numpy/Python version information:

1.16.0.dev0+2ed08ba 3.7.0 (default, Aug 16 2018, 11:04:41)
[Clang 10.0.0 (clang-1000.10.25.5)]

Start by running the provided Python reproducer and compare the arguments passed to array_ufunc for addition and comparison. Trace the scalar arithmetic and comparison dispatch paths to determine where the comparison operand becomes a 0d array. Done means the two operator families have consistent, tested dispatch behavior.