Implementation of (non-nightly) SIMD on FFM

[ggaspersic]

• SIMD operations added to FFM where we do not need the nightly build.
• The biggest trick here is the additional unrolling of the loops, which gives us in total a roughly 15 % reduction on CPU & latency.
• If we use SIMD vectors larger than 4 fields we start to get accuracy issues (hence the additional loop unrolling)