Yoooooo! The reference test is finally deterministic! It finally works!

I'm going to polish this PR tomorrow.

Alright, done. The output of JS code gen is finally deterministic.

The cause

The chain that caused the nondeterministic code gen is as follows:

1. Import IDs are nondeterministic
2. This causes the order in which import adapters are added to be nondeterministic
3. This causes adapter IDs to be nondeterministic
4. This causes the order in which adapters are iterated to be nondeterministic
5. This makes code gen nondeterministic.

So what are import IDs? Import IDs are the arena ID from walrus' ModuleImports. So they are just an index into a memory arena.

Since these IDs are nondeterministic, I suspect that imports are added into the arena in a nondeterministic order. However, I have no idea why that happens. My guess is that walrus just adds them in whatever order they appear in the WASM binary. This would explain why the order seems to depend on compiler versions and OS, and why #4138 (where I made everything in wasm bindgen deterministic) didn't work.

Honestly, I'm not sure what the bug here is. Is the bug that import IDs are nondeterministic, or that wasm bindgen relied on them being deterministic?

The fix

The fix I went with is quite simple: sort imports by name and not ID. Since the non-determinism of import IDs "infects" adapter IDs, I made sure that code gen doesn't depend on their order by sorting imports (and adapters of imports) by import name.

This works, but isn't ideal IMO. I would have preferred to make adapter IDs deterministic, but I couldn't. The wit code handling imports and import IDs is too complex for me.

Again, thank you for figuring this out!
This makes long-term maintenance a bit less annoying, so its very appreciated!