How would you design a production-grade Rust async executor (Tokio-like) from first principles?

A production executor requires a task queue, reactor (I/O driver), worker threads, and a Waker-based notification system.

The executor maintains runnable futures in a work-stealing queue. Worker threads poll tasks. The reactor listens to OS events (epoll/kqueue/IOCP). When I/O completes, it triggers Waker to re-enqueue tasks. The design must avoid contention, ensure fairness, and minimize wake-up overhead.