Experiment with AVX optimizations for FMAC, FMUL operations.

Experiment with AVX optimizations for FMAC, FMUL operations.

Tests how AVX (256-bit operations) fair versus our existing SSE
based optimizations (128-bit). The answer is not well!

vector_math_fmac: unoptimized= 521.3478928421541 runs/ms
vector_math_fmac: sse_unaligned= 512.7810681229649 runs/ms
vector_math_fmac: sse_aligned= 524.1914347119567 runs/ms
vector_math_fmac: avx_unaligned= 468.17904102887024 runs/ms
vector_math_fmac: avx_aligned= 453.4674387705591 runs/ms

vector_math_fmul: unoptimized= 664.2245344616294 runs/ms
vector_math_fmul: sse_unaligned= 666.4911573285701 runs/ms
vector_math_fmul: sse_aligned= 653.6695373653849 runs/ms
vector_math_fmul: avx_unaligned= 613.5270442721115 runs/ms
vector_math_fmul: avx_aligned= 586.0634120611851 runs/ms

FMA operations for FMAC and dot products (used in SincResampler
and WSOLA would probably fare better, but I don't have a machine
which can test these).

BUG=none
TEST=none