Which pattern runs at most one in-flight fetch per cache key in async Rust when many tasks miss at once?