This fixes some Valgrind errors caused by reads from x_by_xplus1 that
used tainted data as an address (see the comments in selfguided_sse4.c
for what's going on).

It also rewrites the algorithm to use an integral image approach
instead of the handwritten filters that the code was using. The end
result is roughly the same efficiency (I think that there's one more
memory load per group of pixels, but this seems not to be measurable)
and I've done some performance optimisation with perf too. Several
32-bit multiplications have been replaced by madd instructions which
do 16-bit multiplications and add adjacent lanes. This is equivalent
to a 32-bit multiplication when the 32-bit lanes contain numbers below
2^15, but runs significantly faster.

Change-Id: I3d0f3043c7861707a56e2fd1849574dc73897d6c