separate SSE and SSE2

[recp]

Separate SSE and SSE2 where SSE2 may not be available e.g i686.

• separate SSE and SSE2
• make Fasth math work both on SSE and SSE2, make -0.0f run on SSE + fast math.
• make some failing test pass when fast math is on
• provide CGLM_FAST_MATH to identify path math
• tests: dont validate nan and inf on fast math

Most operations are SSE floating point but some SSE2 intrinsics are available for instance:

• _mm_castsi128_ps
• _mm_shuffle_epi32

I tried to find some of them and fallback SSE FP domain if possible.

Github action that only works on SSE would awesome actually to identify issues.

READY TO MERGE 🚀

[recp]

Big problem here:

Setting -0.0f is a big issue without SSE2 e.g. SSE + Fasth math which will ignore the sign probably. Also passing 0x80000000 directly to float seems not work.

[recp]

This seems worked:

#if defined(__SSE2__)
#  define GLMM_NEGZEROf ((int)0x80000000) /*  0x80000000 ---> -0.0f  */
#  define GLMM_POSZEROf ((int)0x00000000) /*  0x00000000 ---> +0.0f  */
#else
#  ifdef CGLM_FAST_MATH
     union { int i; float f; } static GLMM_NEGZEROf_TU = { .i = (int)0x80000000 };
#    define GLMM_NEGZEROf GLMM_NEGZEROf_TU.f
#    define GLMM_POSZEROf 0.0f
#  else
#    define GLMM_NEGZEROf -0.0f
#    define GLMM_POSZEROf  0.0f
#  endif
#endif

it defines GLMM_NEGZEROf_TU for each translation unit but better approaches can be considered if available