ggml : move common CPU backend impl to new header

[slaren]

Moves CPU backend declarations in ggml-impl.h to a new header ggml-cpu-impl.h. Avoids unnecessary declarations in other backends that need this header, and workarounds a conflict with nvcc and immintrin.h.

Fixes #9473

[Pleune]

I am still running into this issue when compiling with gcc's -march, I believe due to

llama.cpp/ggml/src/ggml-impl.h

Line 28 in 6dde178

#include <immintrin.h>

Probably the most sane thing to do is not compile with -march... but still wanted to note this.

Edit: to say yes patching a hack in to turn off the __F16C__ flag (replace __F16C with 0), and then adding #include <immintrin.h> to ggml.c has me compiling fine with the broken cuda/gcc combo (I can't easily change the system over to a non-broken set).