Perf: parallelize count_pw_st with OpenMP collapse(2)

[MiniYuanBot]

What's changed?

• source/source_basis/module_pw/pw_distributeg.cpp (count_pw_st):

  • Added OpenMP parallel for collapse(2) to the (ix, iy) double loop for plane-wave stick enumeration
  • Added reduction(+: npwtot_local, nstot_local) for accumulation of total plane-wave and stick counts
  • Added reduction(min/max: ...) for boundary coordinate tracking (lix, rix, liy, riy)
  • This change accelerates the PW initialization stage, which becomes a bottleneck for large FFT grids

• Performance impact (tested on Intel Core i7, GCC 13.3.0, -O3 -fopenmp, grid=256×256×256, repeats=10):

  • Benchmark focuses on the modified count_pw_st function, which is the hotspot in PW initialization for large grids.

Threads	Total (ms)	Avg (ms)	Speedup	Efficiency
1	12363.61	1236.36	1.00	100.0%
2	6111.27	611.13	2.02	101.2%
4	3234.23	323.42	3.82	95.6%
8	2105.93	210.59	5.87	73.4%
12	1851.56	185.16	6.68	55.6%

• Near-linear scaling up to 4 threads (efficiency >95%)

• 8-thread efficiency drops to ~73% due to memory bandwidth saturation

• 12-thread marginal gain diminishes, consistent with SMT overhead on consumer-grade platforms

• Behavior changes: None. The serial code path is preserved when _OPENMP is undefined. All existing MODULE_PW_* unit tests (12/12) continue to pass.

[MiniYuanBot]

\label project_learning
This is Problem 1 of the assignment01 on the plane wave module.
Thanks for the review: )