Fast multiply-add

[ralphtandetzky]

Feature description
If fused-multiply-add (FMA) is supported by the hardware platform, then it should be used.

Describe the solution you'd like
If possible detect at compile-time whether the hardware target supports fused-multiply-add. Use it, if it's there. Otherwise, retreat to a multiply then add implementation, so there won't be a considerable slow-down.

Describe alternatives you've considered
The problem of the above approach is that we'll get different results on different platforms. The behavior is implementation defined. This can be a problem when reproducibility is important. An other approach would be to have the user switch between the alternatives by using compile features.