Speed up scalar Sobol with closed-form evaluation

[wantonsushi]

Hello OpenQMC maintainers!

This PR implements the Ahmed 2024 paper mentioned in issue #67 for the scalar path of sobolReversedIndex. I used the existing benchmark tool to check timings and generate to confirm the output is unchanged.

Changes

• include/oqmc/owen.h: the scalar path uses the closed form.
• src/tools/cli/matrices.cpp: derives the steps from the generator matrices and prints them, the same way it already prints directions[].
• src/tests/owen.cpp: OwenTest.SobolReversedIndex checks the result against the direction matrices for every 16-bit index and dimension.

I also tried to implement a SIMD version (sobol_simd_experiment.cpp). It's not in this PR since it doesn't beat the existing SIMD, but I'm sharing it so you can reproduce the SIMD timings below.

Timings

Using benchmark sobol samples, median of 9 runs, in microseconds:

target	old	new	speedup
CPU scalar	60034	56482	1.06x
GPU (RTX 4070)	128830	124873	1.03x
CPU SSE	47545	62700	0.76x
CPU AVX	47627	62856	0.76x

The speedup is small because generation is only part of a draw, along with scrambling and state hashing. The SIMD implementation fits one dimension per lane (four lanes). The existing SIMD path is faster because it packs all sixteen matrix columns into a wider register, so I left the SIMD paths alone.

Verification

• generate sobol output is byte-identical to before on scalar, SSE and AVX
• Tests pass (168/168), clang-format and clang-tidy clean

[joshbainbridge]

Hi @wantonsushi. I've just given the PR a scan and it is looking good. This is great work, thank you.

Very interesting to see the comparison. I'll follow up with a closer read of the code details sometime tomorrow.