fix(graceful-upgrade): fix uninterruptible grace period and stale FD …

[gabioprisan]

…accumulation

Issue 1 — Uninterruptible grace period after SIGTERM/SIGQUIT: After main_loop consumed signal handlers via tokio::select!, a bare thread::sleep(grace_period) left the process unresponsive to further signals for up to 20 minutes. Replace with a 1-second polling loop that checks an AtomicBool set by a signal-watching task spawned on the still-running server_runtime. A second SIGTERM/SIGINT/SIGQUIT during the grace period now causes an early exit.

Issue 2 — Stale inherited FDs accumulate across upgrade generations: serialize() dumped the entire Fds map, including ports removed from the current config. These zombie FDs were forwarded to every subsequent process generation, holding ports open indefinitely.

Fix with a two-map design in Fds: inherited FDs from deserialize() live exclusively in inherited_pending until a listener claims them via add() or mark_used(). close_unused_inherited() drains inherited_pending, closing stale FDs before send_to_sock() on SIGQUIT. Drop provides a safety net.

Also fix the non-NoSteal listen() path in l4.rs which never called mark_used(), causing active listener sockets to be closed by close_unused_inherited() on third-generation upgrades.

Add integration tests:

• server_phase_gracefulupgrade: SIGQUIT → full phase progression
• server_phase_early_exit: second SIGTERM interrupts grace period