Add avx mixer to test and benchmark
Rework and unroll the avx mixer some more.
The SSE one is 10 times faster than the C one, The AVX is 20 times
faster. The SSE2 function is 5 times faster than the C one.
Let the mixer functions accumulate the intermediate results into a
larger size variable and then clamp to the final precission. This avoids
distortions because of intermediate clamping.
Although the access pattern of the reads are no longer sequential, the
writes are sequential and we don't need to read intermediate values.
Together with the avoided clamping this is probably faster overall.
Add a unit test for the various cases.
