Fix crash running Node-API finalizers during VM destruction

[190n]

What does this PR do?

Split off from #18758.

This will be the fix for #18139 (right now it's incomplete). The issue is that we can run Node-API finalizers after the GlobalObject is destroyed, and the GlobalObject owns the napi_envs, which means that finalizers will read a bunch of bogus data from the space taken up by the napi_env and probably crash (e.g. macOS free() memsets it to zero).

The current way I've tried to fix this is by making napi_envs owned by the ScriptExecutionContext instead of the GlobalObject. But this doesn't work because the ScriptExecutionContext is reference-counted and the GlobalObject holds the last one, so its destructor will free the ScriptExecutionContext and free the napi_envs. A solution could be to make napi_env reference counted and have a ref owned by each finalizer. Then the env just needs to detect if the GlobalObject was already freed when running a finalizer (which it'll be able to do without concern over the napi_env having been freed), and skip it (or we could try running the finalizer anyway, but it would only work if the finalizer avoids using basically any Node-API function).

How did you verify your code works?

There's a test (which currently fails since the fix is incomplete). The test also depends on BUN_DESTRUCT_VM_ON_EXIT from #18758.

[robobun]

^{Updated 5:12 PM PT - Apr 7th, 2025}

❌ @190n, your commit 1fb9a72 has some failures in Build #14605

---

🧪   try this PR locally:

bunx bun-pr 18767