Add f16 vector support

[folkertdev]

closes #484

This is blocked on a proptest release: the required changes have been merged, but not released yet.

[folkertdev]

The latest release of proptest https://crates.io/crates/proptest/1.11.0 includes the f16 feature flag that enables the functionality we need for f16 support here. So assuming CI still passes this is ready.

[folkertdev]

I'm not sure what's up with the wasm tests, locally running wasm-pack test --firefox --headless crates/core_simd works fine but on CI it appears to just get stuck on building the crates.

[calebzulawski]

Might be worth trying a cargo update before trying to actually debug it

[folkertdev]

Turns out this LLVM IR with +simd128 will just get stuck

define <8 x i16> @foo(<8 x half> %x) {
start:
  %1 = fcmp uno <8 x half> %x, zeroinitializer
  %2 = sext <8 x i1> %1 to <8 x i16>
  ret <8 x i16> %2
}

I'll debug that further, of course. Should we block on this? f16 is unstable anyway so it seems unlikely for someone to run into this in practice.

[calebzulawski]

Historically we've identified the function that generates the bad LLVM behavior and replace it with a temporary scalar loop for that architecture. The next best option is if you could find the test that hits it, disable it for that architecture. I don't think we need to block on this, but probably report it to LLVM.

[folkertdev]

I believe it's any comparison on f16, so things like is_finite or is_nan. We don't have a really targeted way of skipping those tests though so we'd just have to skip f16 tests on wasm32 + simd128.

[folkertdev]

@calebzulawski if that's acceptable, this should be ready now. I did open an LLVM issue also llvm/llvm-project#189251.

[folkertdev]

@calebzulawski quick ping, or should I assign someone else?

[calebzulawski]

Sorry about that, thanks!