perf: fix Notify race in writable backpressure

[Yanstart]

Context

Adversarial review found that Notify::notify_one() has a race condition: if fired before the receiver calls notified().await, the notification is lost. The sender then falls back to a 5-second sleep (WRITABLE_FALLBACK_SECS), causing throughput stalls.

Fix (two changes)

1. Replace notify_one() with notify_waiters() for writable notifications in client.rs and server.rs
2. Reduce WRITABLE_FALLBACK_SECS from 5 to 0.1 (100ms) as safety net

Files

• crates/quick-transport/src/client.rs — writable notifier.notify_one() → notify_waiters()
• crates/quick-transport/src/server.rs — same
• bins/quick-entry/src/tunnel.rs — WRITABLE_FALLBACK_SECS 5 → Duration::from_millis(100)
• bins/quick-exit/src/tunnel.rs — same

Acceptance Criteria

• notify_waiters() used for all writable notifications
• Fallback timeout reduced to 100ms
• No 5-second stalls in iperf3 per-second output
• 351+ tests pass