fp::fma should use fmuladd

[gnzlbg]

Currently the floating-point fma vector method performs a fused-multiply add unconditionally, even in targets for which it is slower, or in targets where it is not supported where it has to be emulated (no idea how reliable this emulation is).

It might be better to do a conditional fused-multiply add (llvm.fmuladd) here instead. Those who want to perform a fused-multiply add with extended precision and have its guarantees should probably just use std::arch.