The previous implementation did two separate passes in the horizontal
and vertical directions, with the intermediate values being stored
in a buffer on the stack. This caused bad cache thrashing.

By interleaving the horizontal and vertical passes in combination
with a ring buffer for storing only a few rows at a time the
performance is improved by a significant amount.

Also split the function into 7-tap and 5-tap versions. The latter is
faster and fairly common (always for chroma, sometimes for luma).