ndimage.percentile_filter ignores origin argument for multidimensional inputs

[poolio]

The percentile_filter seems to ignore the origin argument when the input is not a vector. Here is a test example that compares the result of a percentile_filter on a vector vs. an Nx1 array. I would expect the results to be identical, but using non-zero origins yields the same result as using zero origins on the Nx1 array.

import numpy as np
from scipy.ndimage.filters import percentile_filter
x = np.arange(10).reshape(10,1)
y = x.flatten()
x0 = percentile_filter(x, 0, size=(3,1), origin=(0,0)).flatten()
x1 = percentile_filter(x, 0, size=(3,1), origin=(-1,0)).flatten()
y0 = percentile_filter(y, 0, size=3, origin=0)
y1 = percentile_filter(y, 0, size=3, origin=-1)
print 'Error with origin=0:', np.sum((y0-x0)**2)
print 'Error with origin=-1:', np.sum((y1-x1)**2)
print 'Difference between origin=0 and origin=-1 for array:', np.sum((x1 -x0)**2)

Output:

Error with origin=0: 0
Error with origin=-1: 8
Difference between origin=0 and origin=-1 for array: 0

[poolio]

Fixed in #2923