OpenMP Reduce Should do some manual loop unrolling

Currently vtk-m openmp reduction algorithm uses the following approach:

...
VTKM_OPENMP_DIRECTIVE(for schedule(static))
for (vtkm::Id i = numThreads * 2; i < numVals; i++)
{
  accum = f(accum, data[i]);
}

The issue with this approach is that it makes compilers believe that accum is a dependency between each iteration and the loop can't be vectorized. If we look at the assembly for skylake-avx512 generation we see:

.L152:
	vpaddd	(%rcx), %ymm0, %ymm0
	addq	$32, %rcx
	cmpq	%rax, %rcx
	jne	.L152

If we do our own manually unrolling such as:

...
VTKM_OPENMP_DIRECTIVE(for schedule(static))
for (vtkm::Id i = numThreads * 2; i < numVals; i+=4)
{
  auto a = f(data[i], data[i+1]);
  auto b = f(data[i+2], data[i+3]);
  accum = f(accum, f(a,b));
}

We see better vectorization when we look at the assembly for skylake-avx512 generation:

.L152:
	vmovdqu32	(%rax), %ymm1
	vmovdqu32	32(%rax), %ymm9
	vmovdqu32	64(%rax), %ymm4
	vmovdqu32	(%rdi), %ymm0
	vmovdqu32	64(%rdi), %ymm5
	vmovdqu32	96(%rax), %ymm8
	vmovdqa64	%ymm1, %ymm3
	vpermt2d	96(%rdi), %ymm2, %ymm5
	vpermt2d	%ymm9, %ymm2, %ymm3
	vmovdqa64	%ymm4, %ymm7
	vpermt2d	32(%rdi), %ymm2, %ymm0
	vpermt2d	%ymm8, %ymm2, %ymm7
	vpermt2d	%ymm5, %ymm2, %ymm0
	vpermt2d	%ymm9, %ymm6, %ymm1
	vmovdqa64	%ymm3, %ymm5
	vpermt2d	%ymm8, %ymm6, %ymm4
	vpermt2d	%ymm7, %ymm2, %ymm5
	vpermt2d	%ymm4, %ymm2, %ymm1
	vpermt2d	%ymm7, %ymm6, %ymm3
	vpaddd	%ymm5, %ymm0, %ymm0
	vpaddd	%ymm3, %ymm1, %ymm1
	vpaddd	%ymm1, %ymm0, %ymm0
	subq	$-128, %rax
	vpaddd	%ymm0, %ymm10, %ymm10
	subq	$-128, %rdi
	cmpq	%rdx, %rax
	jne	.L152