LHCbMath - Replace runtime CPU dispatching with compile time implementation selection.

The following discussion from !2265 (merged) should be addressed:

• @clemenci started a discussion: (+3 comments)

  I'm not convinced this it the right approach... I mean we still do dynamic dispatching, but we aim for the platform we build for instead of what we find at runtime.

  Cannot we just build only the sse4_2 version for the base platform and build only the avx2 version for avx2 builds? (if available of course). I thought we were going to replace the runtime dispatch by a compile time one.