perf: replace threads 100ms polling loop with poison-pill shutdown

Parent issue

Part of #156 (performance audit findings).

Problem

threads/actor.rs uses rx.recv_timeout(Duration::from_millis(100)) in the message loop to periodically check cancellation_token.is_cancelled():

match rx.recv_timeout(Duration::from_millis(100)) {
    Ok(mut envelope) => { /* handle message */ }
    Err(RecvTimeoutError::Timeout) => {
        if cancellation_token.is_cancelled() { break; }
    }
    Err(RecvTimeoutError::Disconnected) => break,
}

This causes:

• Up to 100ms shutdown latency after stop() is called
• 10 wakeups/sec per idle actor (constant CPU churn)
• Unnecessary overhead for applications with many idle actors

Proposed fix

Replace the timeout-based polling with a poison-pill approach:

1. Wrap mailbox messages: enum MailboxItem<A> { Message(Box<dyn Envelope<A>>), Shutdown }
2. When stop() is called, send MailboxItem::Shutdown through the channel
3. Replace recv_timeout with blocking rx.recv()
4. On Shutdown, break the loop — instant cancellation, zero polling

This keeps CancellationToken for timers/streams but removes the polling loop from the core message path.

Files to change

• concurrency/src/threads/actor.rs — message loop, stop() impl
• Possibly rt/src/threads/mod.rs — if channel type changes

Testing

• Existing thread actor tests should pass unchanged
• Add test verifying shutdown latency < 10ms
• Add test verifying idle actor has no CPU wakeups

Start in concurrency/src/threads/actor.rs by tracing the message loop and stop() implementation, then check rt/src/threads/mod.rs if the channel type crosses that boundary. Review the existing thread actor tests before changing the mailbox flow. Done means shutdown no longer relies on recv_timeout polling, and the existing tests plus latency and idle-actor wakeup tests pass.