Faster OpenMP Reorder and Accumulation Kernels

[kaanolgu]

Speeds up the OpenMP reorder and sum_intox kernels by removing the per element index computation. The transpose used to call get_index_reordering for every element, which nests two more routines doing integer mod and division per element and blocks vectorisation

This addresses the same bottleneck as #131 (related to #78), where reorders and accumulations were measured at roughly half of the transeq runtime while the per element path sustains only 10 to 20% of peak bandwidth. #131 specialises the hot transeq mappings with explicit cache blocked kernels. This PR takes an alternative, more general route: the index map advances by a fixed stride along the leading index, so the kernels compute a base and stride once per column and run a plain inner loop. Because the same routine covers every directional mapping rather than a selected few, the speedup extends across the divergence and gradient reorders as well as transeq, broadening the gain over the full step while keeping the implementation compact

Padded dimensions are exact multiples of SZ, so the leading index stays inside one SZ block and never crosses a mod or division boundary. One output index then advances by a fixed stride while the other two stay constant. Evaluating the map at i = 1 and i = 2 recovers the base and stride, and the inner loop reproduces the original writes exactly. This needs dims(1) == SZ (a directional field). A DIR_C field has a leading dimension of nx_padded, so the kernels guard on dims(1) == SZ and fall back to the per element loop there

Performance (TGV 256^3, 100 iterations, generic 2decomp FFT)

1 rank, 24 threads:

phase	before (s)	after (s)	change
transeq	22.0	17.3	-21%
divergence	10.0	7.32	-27%
gradient	9.5	6.79	-28%
poisson FFT	113	111	(untouched)

8 ranks, 3 threads each:

phase	before (s)	after (s)	change
transeq	19.1	15.35	-20%
divergence	8.55	6.22	-27%
gradient	7.8	5.68	-27%
poisson FFT	18.6	18.44	(untouched)
avg step	0.576	0.508	-12%

At one rank the generic FFT is about 74% of the step so the total barely moves. At eight ranks the FFT has scaled down and the saving becomes a 12% step reduction, so the benefit grows with rank count

[ia267]

LGTM!