This makes SSE2 version match the (correctly-rounded)
scalar reference (which the current version doesn't:
see https://godbolt.org/z/hK5938jda).
The new implementation uses the same idea as mentioned in
https://github.com/hrydgard/ppsspp/pull/20870#issuecomment-3413134437
It may be a bit slower than current implementation (which
simply truncates?), but correctness/consistency is
probably worth it.
The difference is slight (by 1 in RGB color channels),
but observable in synthetic examples (in particular the
current implementation darkens completely opaque parts,
including `0xFFFFFFFF` -> `0xFFFEFEFE`, and, more
strikingly,`0xFF010101` -> `0xFF000000`).
This doesn't touch NEON part at all (so "vibe-coded" comment
stays for now), but it shouldn't be too hard to port.
fp64
9932ad3f85