Improved wasm benchmark support and add support for wasm +relaxed-simd

[bluss]

Add support for +relaxed-simd in wasm, which increases performance and should degrade gracefully when not available. relaxed-simd adds a fused multiply-add instruction, except that it doesn't guarantee fused operation; it can also be implemented by multiply then add (which leaves us where we started).

• Add short contributing document
• Support --wasm in the benchmark runner, selectable features (simd128 with or without relaxed)
• Add justfile to simplify running tests through wasmtime

Small refactoring in the wasm loop.

Unrolling the loop by 2 was explored - with just +simd128 it is a negative, with +simd128,+relaxed-simd it helps a bit on my intel laptop and helps nothing on my apple m4 laptop. Decision: no unrolling for now, keep it simple.

[byeongjee]

Sounds good! I'll take a look tonight

[byeongjee]

@bluss With +simd128 and without +relaxed-simd, there was about a 6.5% degradation on my M2 Pro Mac, compared to the current master branch. I guess there might be some microarchitectural differences between M2 and M4, making loop unrolling harmful in M2.

When I unrolled the loop only if +relaxed-simd is on, I could avoid this (at the cost of longer, duplicated code).

If supporting a byte-precise matmul variant is also a main goal (so people would use +simd128 without +relaxed-simd), you could try that?

[bluss]

Thanks for testing! I don't have any goals myself for wasm, so it's actually better if you as a user decide this. Is your use case only for +simd128 so far? (+relaxed-simd seems to be held back by lack of Safari support?)

Changing unrolling only for +relaxed-simd seems like the only way forward (or not changing it at all).

[byeongjee]

Is your use case only for +simd128 so far?

Not really so far, but I think it's good to change unrolling only for +relaxed-simd if it's not a big burden.
Thanks!

[bluss]

Decision is to skip unrolling since it has no benefit on my m4 system (I benchmarked the sensitivity to that now), keep it simple. Now the PR just remains as a benchmark tool improvement, refactoring and opt-in relaxed-simd benefit.