stdlib: implement first rounding methods for `Float32` and `Float64`

[soc]

See https://github.com/dinfuehr/dora/issues/140.

[soc]

@dinfuehr could you have a look?

[dinfuehr]

Why would we want to add e.g AVX encoding as well? I suppose advantages (if there are any) are miniscule at best but adds one more configuration to Dora.

[soc]

Hi @dinfuehr, this is due to https://github.com/dinfuehr/dora/issues/267. It's not that terrible, it's basically another line here:

   static ref HAS_ROUND: bool = is_x86_feature_detected!("sse4.1");
++ static ref HAS_AVX: bool = is_x86_feature_detected!("avx");
   static ref HAS_POPCNT: bool = is_x86_feature_detected!("popcnt");
   static ref HAS_LZCNT: bool =  is_x86_feature_detected!("lzcnt");
   static ref HAS_TZCNT: bool = is_x86_feature_detected!("bmi1");

I would have preferred only using AVX to start with, but requiring AVX is probably a few years too early (and wouldn't make sense with all the other checks we do here).

[dinfuehr]

We can add that later but let's not add all these optimizations which atm just complicate the code but do not bring any real speedup.

[soc]

Ok!

[soc]

@dinfuehr CI fails seem to be unrelated, I think one of your refactorings broke the linter.

[dinfuehr]

Thanks, should be fixed now.

[soc]

@dinfuehr all green!

[soc]

@dinfuehr I think this should be good to go, the CI failure seems to be unrelated.

[soc]

@dinfuehr could you have a look?

[soc]

@dinfuehr ping?