feat: OpenMP + AVX-512/AVX2 parallelization for multi-core performance

[aminkvh]

Summary

This PR adds OpenMP multi-threading and AVX-512/AVX2 SIMD vectorization to
FreeSASA, giving 5-8× speedups on multi-core machines with no changes to the
public API. All 54 unit tests pass with bit-identical results.

What Changed

• src/sasa_lr.c — Replaced pthreads with OpenMP; removed the hard
  thread cap; added dynamic scheduling
• src/sasa_sr.c — AVX-512/AVX2 SIMD inner distance check using a
  float32 SoA neighbor cache; scalar/auto-vectorization fallback for
  portability; C99-compatible aligned allocation (posix_memalign wrapper)
• src/nb.c — Two-phase parallel neighbor-list construction (was
  entirely serial O(N²))
• src/freesasa.c — New freesasa_calc_structures_parallel() for
  processing multiple structures (trajectory frames) concurrently
• src/freesasa.h — Declaration for the above; no existing API changed
• src/main.cc — CLI auto-detects core count via omp_get_max_threads()
• CMakeLists.txt — CMake build system (existing autotools build unchanged)

Performance (58,674-atom structure)

Algorithm	1 thread	8 threads	16 threads
L&R standard (20 slices)	1.19 s	0.23 s	0.18 s (6.5×)
L&R high-res (100 slices)	4.08 s	0.62 s	0.48 s (8.4×)
S&R (100 pts, AVX-512)	0.27 s	0.13 s	0.13 s (2.1×)

Backward Compatibility

• All 54 existing unit tests pass with bit-identical results
• OMP_NUM_THREADS=1 or parameters.n_threads=1 restores single-threaded behavior
• No existing API function was modified or removed
• FREESASA_DEF_NUMBER_THREADS is preserved and exported correctly
• C99 standard maintained throughout (no C11 APIs used)

Notes

I noticed in CONTRIBUTING.md that new features require unit tests. The new
freesasa_calc_structures_parallel() function is currently covered by the
Python test suite. I am happy to add a pure C unit test if you would prefer
that before merging.