Add round

[RunDevelopment]

See #355

This implements round for all floating-point types to match std::simd.

On ARM, it gets lowered to the vrndaq family of instructions. On x86 and WASM it's emulated using trunc. The fallback uses scalar f<N>::round as usual. Because of the fallback, we now also depend on libm's round{,f} functions.

Notes on the emulation:

• It's the same trick LLVM uses, so it naturally gives bit-exact results.
• My use of copysign(a, b) might not be perfectly optimal. On x86 and WASM, it's implemented using bitwise masking. Part of that is setting the sign bit of a to 0. The emulation uses a positive constant so that part is unnecessary. I'm not sure if LLVM can optimize it away.
  Update: LLVM successfully optimizes copysign to use VPTERNLOG with AVX512. So emulation is optimal with AVX512.

[Shnatsel]

Looks good. Thank you!