fix weak memory bug in TLS on Windows

library/std/src/sys/pal/windows/thread_local_key.rs

@@ -141,9 +141,15 @@ impl StaticKey {
  panic!("out of TLS indexes");
  }
 
- self.key.store(key + 1, Release);
  register_dtor(self);
 
+ // Release-storing the key needs to be the last thing we do.
+ // This is because in `fn key()`, other threads will do an acquire load of the key,
+ // and if that sees this write then it will entirely bypass the `InitOnce`. We thus
+ // need to establish synchronization through `key`. In particular that acquire load
+ // must happen-after the register_dtor above, to ensure the dtor actually runs!
+ self.key.store(key + 1, Release);
+
  let r = c::InitOnceComplete(self.once.get(), 0, ptr::null_mut());
  debug_assert_eq!(r, c::TRUE);