Optimize GintAtom::set_phi() in module_gint

[chaomianpkp]

Summary

Optimize GintAtom::set_phi() in module_gint, which is a hotspot on the
cal_gint_rho / cal_gint_vl path in LCAO grid integration.

Changes:

• Refactor single-meshgrid logic into set_phi_at_meshgrid()
• Micro-optimizations (enabled by default): compute coord.norm() once,
  use inv_dist instead of repeated divisions, cache rcut/nw/nwl/dr_uniform
• Optional CMake flags: GINT_SET_PHI_SIMD, GINT_SET_PHI_OPENMP
  (nested OpenMP in SCF is intentionally disabled when already inside a parallel region)

Motivation

set_phi() is called for every atom on every fine meshgrid inside each BigGrid.
The original implementation repeated norm/division and loop-invariant lookups.

Test plan

• Build with ENABLE_MPI=OFF, USE_OPENMP=ON, ENABLE_LCAO=ON
• Build with -DGINT_SET_PHI_SIMD=ON
• ctest -R gint (if BUILD_TESTING=ON)
• SCF on tests/performance/P110_C2H6O_lcao (ethanol + vacuum box)
  • Correctness: etot / stress unchanged within floating-point tolerance (~1e-10)
  • Performance (8 OpenMP threads, serial binary):
    • cal_gint_vl: 5.70s → 1.86s (~3.1×)
    • cal_gint_rho: 5.29s → 1.53s (~3.5×)
  • #pragma omp simd on the iw loop shows no measurable gain on tested platform

Notes

• Does not change numerical formulas; only code structure and local efficiency.
• GINT_SET_PHI_OPENMP is for non-nested contexts only; normal SCF uses micro-opts only.