fix(media): break OutboundClockSync Arc cycle leaking ~70 KB/call

[ftong2010]

Problem

Under sustained load (50 cps), rustpbx memory grows without bound — roughly
400–780 MB/min, reaching the 8 GiB container limit in ~10 minutes. About
70 KB per call is allocated and never freed.

Before fix

After fix

Root cause

Memory Profiling point to session building process strongly. With step by step metrics collection and analyzing, root cause is identified

Commit 5f4a3106 ("fix(media): stop WebRTC Opus IVR pops and harden egress
timeline", 2026-08-21, jinti) added the OutboundClockSync RTP observer and
registered it on the PeerConnection with a strong pc reference:

pc.rtp_observers → OutboundClockSync → pc      (cycle #1)

That alone keeps the PeerConnection alive. Even after replacing pc with
the RtpSender, the observer is also attached to the RTP transport
(attach_registered_observers), and the sender holds that same transport:

transport.observers → OutboundClockSync → sender → sender.transport = transport   (cycle #2)

Rust reference counting never collects either cycle, so the RtpTransport
and its IngressTap observer (also registered on the transport) are retained
forever.

Fix

Hold a Weak<RtpSender> in OutboundClockSync instead of a strong
PeerConnection/RtpSender, upgrading on use in on_egress. A Weak does
not keep the sender alive, so once the PeerConnectionInner drops, the sender
drops, the transport's last strong ref goes, and the transport + tap free.

Verification

• Instrumented RtpTransport/IngressTap lifecycle counters: before the fix
  3600 created / 0 dropped; after, 3600 / 3600.
• jemalloc stats.allocated stays flat across repeated 1,800-call runs
  (~62 MB) instead of growing ~70 KB/call.
• 50 cps × 10 min with recording + CDR + RWI enabled: memory plateaus at
  ~470 MiB (was unbounded).