Don't multiply individual pixels by combined row/column weights.
Instead, calculate the columns separately, using only the row
weights, and then combine them using the column weights.  This
improves accuracy sufficiently so that when all the input value
are 1, the result is not greater than 1, which could previously
happen (or at least it's far less likely to happen now.)

Performance should be at most marginally worse for 4-component
pixels, about the same for 3-component pixels, and slightly better
for 2- and 1-component pixels.