Optimize round and roundf

[bvinc]

In reference to #154, this optimizes the round and roundf functions. This does not change the behavior at all.

I compared all f32 values to verify that the behavior of roundf has not changed.

For round, I directly translated the roundf function and verified that it behaved the same with random bits for the sign and exponent and beginning and end of the mantissa.

I ran some criterion benchmarks and basically both functions went from about 7ns to about 3ns.

round/round             time:   [7.1414 ns 7.1470 ns 7.1543 ns]                         
round/round_new         time:   [3.0178 ns 3.0311 ns 3.0471 ns]                             

roundf/roundf           time:   [6.8636 ns 6.8767 ns 6.9016 ns]                           
roundf/roundf_new       time:   [2.8326 ns 2.8401 ns 2.8562 ns]                               

[Lokathor]

Looks like we had some CI trouble, but actually we haven't been passing CI lately to begin with.

It appears that none of the failures were because of thia PR.

My only thought is that in general we might want to mark these and other functions as inline, but that can be done later.

@bvinc I can merge if you're all done.

[bvinc]

I'm done, but I'll add #inline too if you want. Looks like it should definitely be on copysign/copysignf. Are you sure we also want inline on trunc/truncf?

[Lokathor]

oh i forgot how big that one is.

we'll do inline separate, I'll just merge for now.