Use muladd in complex evalpoly

[augustt198]

Speeds up evalpoly for complex inputs a bit by using muladd for both recurrences. The number of instructions generated by the inner loop is halved.

[MasonProtter]

Looks like the old version had a comment on it regarding this issue which has since been closed.

[StefanKarpinski]

I'll leave this for a bit longer in case @stevengj has any comments. Please ping in a bit if this hasn't been merged and I'll merge it.

[stevengj]

Looks good to me. LLVM seems to emit a fused multiply-subtract these days:

julia> f(a,b,c) = muladd(-a, b, c)
f (generic function with 1 method)

julia> @code_llvm f(1., 2., 3.)

;  @ REPL[1]:1 within `f'
define double @julia_f_16811(double, double, double) {
top:
; ┌ @ float.jl:410 within `muladd'
   %3 = fmul contract double %0, %1
   %4 = fsub contract double %2, %3
; └
  ret double %4
}

julia> @code_native f(1., 2., 3.)
	.section	__TEXT,__text,regular,pure_instructions
; ┌ @ REPL[1]:1 within `f'
; │┌ @ REPL[1]:1 within `muladd'
	vfnmadd213sd	%xmm2, %xmm1, %xmm0
; │└
	retq
	nopw	%cs:(%rax,%rax)
; └