Remove handcrafted MMX code

[mstembera]

This removes the handcrafted MMX code as discussed here #4786 (comment)

It also fixes a Makefile bug (line 157) where we claimed sse support for x86-32 but didn't actually enable it.

Finally I tried making it more explicit that we have two different kinds of "architecture support". In some cases we have custom handcrafted code using intrinsics. In other cases we simply enable the appropriate -m compiler flag allowing the compiler to emit advanced instructions on its own.

As @vondele pointed out maybe some CI stuff should be looked at.

No functional change
bench: 1246812

[Fanael]

vec_cleanup in src/nnue/nnue_feature_transformer.h should be removed entirely as well, it was only ever needed for MMX, all other vector ISAs have empty implementation.

(At assembly level, AVX does require cleanup as well, but the compiler takes care of inserting vzerouppers for us in function epilogues where necessary, so we don't have to worry about it at source level.)

---

It also fixes a Makefile bug (line 157) where we claimed sse support for x86-32 but didn't actually enable it.

That was done deliberately in commit e178a09, the corresponding messages were just not updated.

[mstembera]

@Fanael Thank you for your comments. I have updated the PR based on them.