Loop inside SIMD (instead of in the caller) so that we can handle
multiple 4px blocks per iteration, allowing for more efficient
SIMD. To make this easier, also transpose the masks for the hor
filter.