Fix #6286: avoid redundant libxc init_func calls in per-point evaluation paths

[mohanchen]

Fix #6286
Description:

Issue #6286 reported that after SCF convergence, XC_Functional_Libxc::init_func was called over 3000 times while running the 08_EXX/01_GO_S1_HSE test.

Root cause: the per-point libxc wrappers (xc_spin_libxc, gcxc_libxc, gcxc_spin_libxc, tau_xc, tau_xc_spin) called init_func -> evaluate on a single grid point -> finish_func inside the grid-point OpenMP loop. init_func only depends on the functional id, polarization, and runtime parameters (hybrid_alpha, hse_omega), none of which vary across grid points, so it needs to run only once per thread rather than once per point.

Changes:

1. Added overloads of all five per-point wrappers that accept an already-initialized std::vector<xc_func_type>. The original func_id overloads are kept as thin wrappers that init, delegate, and finish, so existing callers and tests are unaffected.
2. In gradcorr_xc_kernel, initialize the functional once per OpenMP thread (inside #pragma omp parallel, before the for loop) and reuse it for all grid points handled by that thread.
3. In v_xc's nspin==4 branch, split #pragma omp parallel for into a parallel block that initializes the functional once per thread followed by the for loop.
4. Added a std::atomic counter in init_func with get_and_reset_init_count(), exposed via libxc_abacus.h, so unit tests can assert that per-point paths do not re-initialize the functional.
5. Added unit tests verifying that the new funcs overloads produce bit-identical results to the func_id overloads for LDA, GGA, and mGGA, and that the init count is exactly 1 for the old overload and 0 for the new one.
   Result: init_func calls on the per-point paths drop from O(nrxx * SCF steps) to O(nthreads * SCF steps), eliminating the 3000+ calls observed in the issue. Physical results are unchanged because the functional evaluation itself is identical.

Verification:

• Built with ENABLE_LIBXC=ON (libxc 7.0.0): make -j 30 exits with 0.
• All 10 XC test executables pass: VXC (18), GRADCORR (3), PBE (23), PZ_SPN (14), SCANL_LAPL (5), HSE (9), SCAN (1), LAPL (4), NCGGA_RADIAL (9), NCGGA_DISCRETE_FD (40).
• code_quality_score: 8/8 changed files pass (average 80.6, minimum 61, pass line 60).
  Note: this PR addresses only the per-point (A-class) redundant init_func calls. The remaining per-SCF-step init_func calls in v_xc_libxc and related whole-grid functions (B-class) are left for a follow-up, where a cached RAII functional handle can reduce them to once per calculation.