Because we have an (effective) 3-pixel border around each
processing unit, and the local sums in the self-guided filter are
only taken over at most 5x5 regions, we have 1 pixel's worth of
spare border.

We can use this border to greatly simplify the filter: Instead
of calculating a 64x64 region of the A[] and B[] arrays, we can
calculate a 66x66 region. Then we don't have to deal with complicated
boundary conditions when generating the final 64x64 output block.

This also makes a few other related changes:
* The 'boxnum' function has been effectively redundant
  for a while - due to the way we do the 5x5 (or 3x3) windowing,
  the values we actually use are always (2r+1)^2. So we can skip
  calling this function if MAX_RADIUS <= 2

* We can remove the annoying special case for tiny processing units
  in the self-guided filter, as we no longer have to worry about
  border behaviour

* We change the SSE4.1 code to match the new C code, removing a ton
  of complexity. Further refactoring/speedups are probably
  now possible, but this includes the minimal changes to pass all
  the tests.

Change-Id: I99beee164a31349a5228a9bef048e5f35c9639f2