Optimize float<->integer conversions further

[Shnatsel]

#310 which adds 64-bit conversions left some performance on the table

also, the 32-bit AVX2 conversions have a branch inside them, which might cause rampant misprediction; looking for branchless formulations may be worthwhile

there's also a mention of a "magical mystery algorithm taken from LLVM" for SSE and we might want to plumb LLVM for more of those

[valadaptive]

also, the 32-bit AVX2 conversions have a branch inside them, which might cause rampant misprediction; looking for branchless formulations may be worthwhile

Did you actually benchmark this? If you're referring to the ones I added in #134, the branch is only hit if the operands are huge (above i32::MAX). The intent is that this should be extremely rare, even for f32->u32 where the range is technically valid. I already addressed a review comment in that PR to the same effect.

[Shnatsel]

Yes, I didn't look very closely at the time of writing the issue but it does appear that the branch is very rare and should be beneficial to performance overall.

Despite its rarity, outlining it into a #[cold] function likely isn't worth it, based on my benchmarks of #323