Reduce methods give wrong output with shapes that contain 0

[kwgoodman]

The reduce methods (sum, mean, max, lastrank, etc) do not return the correct output when the input is ndim > 1 with a shape that contains a zero.

For example, la.larry(np.zeros((2,0,3))).sum(1) returns NaN but should return:

>> la.larry(np.zeros((2,0,3))).sum(1)
label_0
    0
    1
label_1
    0
    1
    2
x
array([[ NaN, NaN, NaN],
       [ NaN, NaN, NaN]])

This bug is related to Bug #518442.

which would match numpy:

>> np.nansum(np.zeros((2,0,3)), 1)

array([[ NaN, NaN, NaN],
       [ NaN, NaN, NaN]])