Use `core::arch` intrinsics for for `rint` and `rintf` where possible

[hanna-kruppe]

These functions don't have any arch-specialized versions right now. I believe they can be done with smaller code size and better performance on some targets:

• On Wasm, f{32,64}_nearest are always available and already exposed in core::arch (cc Use core::arch::wasm functions rather than intrinsics #418).
• AArch64 with target_feature=neon has frintn but the scalar versions don't seem to be exposed in core::arch. I guess the vector versions (e.g., vrndn_f32) are still an improvement over softfloat even if they need an unnecessary scalar->vector splat.
• On x86 and x86_64, there's _mm_round_{sd,ss} and _MM_FROUND_TO_NEAREST_INT. But this requires SSE4.1 so it's not applicable unless someone happens to build libm from crates.io with appropriate RUSTFLAGS.
  • GCC manages to emit inline code for it with just SSE2 but it's not exactly pretty and I'm not sure if it's even significantly better than the generic implementation in this crate.

[tgross35]

This would be very welcome, are you able to put up a PR? They just need to be exposed in math/arch and then select_implementation! in the relevant functions.

If you aren’t able to, I’ll try to get to this at some point.

[hanna-kruppe]

I have very limited time for non-day job contributions, I think I'll be able to cobble together something from the godbolt experiments I did while writing this issue but if it gets more complicated I'll have to bail out and leave the rest to you.