Fix subleaf in Intel CPUID

[AntoinePrv]

The original behaviour was maintained during the CPUID refactor.
The subleaf is ignored so the the behaviour must be incorrect.
I believe Intel compiler supports __cpuidex but I am not sure.

[AntoinePrv]

@serge-sans-paille this is ready

[serge-sans-paille]

It's actually supported only by decently recent intel compiler, but also by gcc and clang with the most recent version and providing <cpuid.h> is included. I've shared the minimal version through this link:
https://godbolt.org/z/jjx8z9dsh

[AntoinePrv]

Alternatively, if we're going to have an assembly version anyways, should we always use it?

[serge-sans-paille]

I think we can rely on the asm one, less code to maintain. Just double check that the assembly is actually as good as the one from __cpuidex

[AntoinePrv]

They are very similar, I am am not knowledgeable enough to tell how it matters.
Though if I understand correctly, CPUID is typically ~100 clock cycles, and the result is cached, so this should not be what matters.

godbolt

[AntoinePrv]

I can push the ASM fix, or we can also SFINAE detect if __cpuidex is available. Let me know what you think is best.

[serge-sans-paille]

Let's keep only the asm, and the slight difference of clang saving one instruction compared to icc and gcc is negligible compared to the invocation cost of cpuid. Just keep a comment mentioning __cpuidex and why we just use the asm (same code in the end, less portability issues)