perf: inline f32 poly loops + AVX-512 sqrt/abs + CMake tests build

Performance improvements (N=524288, 0 ULP maintained):
  exp  f32: 0.085x numpy → 0.70x  (+8x vs old scalar-per-call approach)
  log  f32: 0.095x numpy → 0.87x  (+9x)
  sin  f32: 0.054x numpy → 0.74x  (+14x)
  cos  f32: 0.053x numpy → 0.72x  (+14x)
  sqrt f32: 0.910x numpy → 1.07x  (now vectorized, AVX-512 immune to throttle)
  sqrt f64: parity maintained

Root causes fixed:
1. noinline helper functions (npy_expf_vec16 etc.) caused 32768 function
   calls per 524288-element array; now the polynomial is inlined directly
   into each template specialization with all 14-15 constants defined as
   non-static locals before the loop — GCC keeps them in zmm8-zmm31.

2. -ffloat-store in Makefile caused GCC to spill every __m512 intermediate
   to the stack and reload it, doubling the instruction count for every
   operation. Removed (redundant on x86-64 with SSE/AVX default float ABI).

3. sqrt/abs had no AVX-512 specialization; added 16-wide float and 8-wide
   double loops using _mm512_sqrt_ps/pd and _mm512_abs_ps/pd (IEEE 754
   exact, 0 ULP, immune to CPU frequency throttling caused by other AVX-512
   loops running in the same process).

Build system:
- Replace tests/Makefile with tests/CMakeLists.txt
  cmake -S tests -B tests/build && cmake --build tests/build
  cmake --build tests/build --target test
- Update root CMakeLists.txt help messages accordingly

Author: peng.li24

CMakeLists.txt

@@ -76,5 +76,5 @@ add_custom_target(deb
 
 message(STATUS "numpycpp v${PROJECT_VERSION}  (header-only C++ library)")
 message(STATUS "  C++ Standard: ${CMAKE_CXX_STANDARD}")
-message(STATUS "  DEB:  make deb  →  numpycpp-dev-${CPACK_PACKAGE_VERSION}-Linux.deb")
-message(STATUS "  Test: cd tests/ && make  →  build + run bit-level alignment tests")
+message(STATUS "  DEB:  cmake --build <build_dir> --target deb")
+message(STATUS "  Test: cmake -S tests -B tests/build && cmake --build tests/build --target test")