Remove `msub` operation

[ajakubowicz-canva]

Context: #14 (comment)

[LaurenzV]

First we should verify that the negating thing indeed leads to efficient code generation.

[awxkee]

FWIW LLVM can recognize all kinds of FMA variants starting from Rust 1.45 ( for x86_64 ):

https://godbolt.org/z/f3G4WEM6P

https://godbolt.org/z/61jYsxr83

[raphlinus]

That matches my experiment with achieving vfms on Neon: https://rust.godbolt.org/z/YTe4bcz4n

Maybe there's more validation to be done, but from my perspective I think we can now confidently say that LLVM is able to optimize this particular class of intrinsics. Much less so for other things I'm concerned about, like round/convert combinations, so I think more investigation is needed on those.