Added a thread_local guard to trigger dtor hook before fibers are used

[ohadravid]

Follow to rust-lang/rust#148799, specifically @alexcrichton's comment.

Summary: rust-lang/rust#148799 switches the way destructors for thread locals are scheduled on Windows, and will now use Fiber Local Storage (FlsAlloc accepts a callback that is called on fiber/thread/process exit).

Because wasmtime uses Fibers on Windows to support async functions, it now need to make sure that this hook is registered before Fibers are used.

To do that, we create an additional TLS variable with a Drop impl, and use it just before calling ConvertThreadToFiber. Because ConvertFiberToThread is already called before exiting, this guarantees that the hook will both be registered correctly, and invoked when the thread exits.

If a fiber does exit abnormally, it's possible that ConvertFiberToThread won't be called, in which case thread locals with Drop impls will be leaked.

@alexcrichton - let me know if you want me to add a resume_separate_thread-like test that uses thread locals + a "native" os thread (== one that isn't initialized with the Rust runtime).

[alexcrichton]

Thanks for this! If you've got a test that'd be awesome to add, but no need to go out of your way for that. You can feel free to add it to this crate itself or in the top level tests/* as appropriate