Use fused multiply-add instructions when available

[JSorngard]

This PR makes the polynomial function use fused multiply-add instructions when the target CPU supports them. This speeds up function execution slightly in those cases.

[JSorngard]

This means some tests started failing, because the exact errors have been shifted around by the difference in round-off error.

As a result they were made slightly less stringent.

[JSorngard]

In the original Fortran code by Fukushima the compiler is allowed to pick which instructions to use. E.g. if the user sets "-march=native" it may use fused multiply-add instructions. As a result their use shouldn't conflict with how the algorithm is set up. Though it is possible that the minimax algorithm that determined the coefficients for the polynomials would have determined different coefficients if it ran with fma instrucitons (I don't know if it did).

[JSorngard]

I will need to determine if the fma instructions break the minimax approximation before I merge this.