fix: end the session when a UAS 2xx to an INVITE is never ACKed

[tgeorge06]

Problem

When a server INVITE transaction enters Completed it arms Timer K = T4 (5 s by default). Timer K terminates the transaction well before 64·T1 (32 s), which causes two problems:

• A 2xx that is not ACKed is retransmitted for about 5 s instead of 32 s. A non-2xx likewise stops at T4 instead of Timer H.
• When the ACK never comes, the dialog is stuck. It stays in WaitAck (initial INVITE) or Confirmed (re-INVITE) forever, with no state event and no BYE. The peer may consider the call up while no media flows, and the application cannot see that anything is wrong.

In addition, Timer G doubled up to 64·T1 rather than T2, so later retransmissions were 8 s and 16 s apart.

Reproduction

New src/dialog/tests/test_uas_ack_timeout.rs uses a raw UAC (UDP, and TCP for the reliable-transport case) and the usual UAS loop. Timers are short but keep the RFC's T4/T1 ratio: T1 = 20 ms, T4 = 200 ms, 64·T1 = 1.28 s.

• No ACK for the initial INVITE: the 2xx must be retransmitted with growing intervals until the last part of 64·T1. The dialog must then emit Terminated(Timeout), and a BYE must arrive no earlier than 64·T1, with the right Call-ID, From tag (the 2xx's To tag) and To tag.
• T2 cap (test_2xx_retransmission_interval_doubles_up_to_t2): with t2 = 4·T1, the Timer G durations the transaction schedules double from T1 and stay at or below T2. It checks the scheduled durations rather than wall-clock arrival times, which a loaded machine stretches.
• Reliable transport (test_2xx_over_tcp_is_retransmitted_until_the_ack): over TCP the 2xx is retransmitted until the ACK arrives. On main it is sent once (got 1 transmissions), so with the teardown below a single 2xx lost behind a later hop would end the call.
• No ACK for a re-INVITE on an established call: the same retransmission, Terminated(Timeout) and BYE.
• Control: when the ACK arrives, retransmissions stop, no BYE is sent and the dialog stays Confirmed.

Fix

• transaction.rs: a server INVITE entering Completed no longer arms Timer K. Timer D (64·T1) already ends the transaction; for a non-2xx it acts as Timer H, and for a 2xx it is the retransmission limit. Confirmed still arms T4 (Timer I).
• transaction.rs: Timer G doubles up to the new EndpointOption::t2, which defaults to 4 s (the RFC value).
• transaction.rs: Timer G is also started for a 2xx on reliable transports (TCP/TLS/WS). RFC 3261 §13.3.1.4 has the UAS core retransmit the 2xx on every transport, because it can still be lost at a later UDP hop. A non-2xx is still retransmitted on unreliable transports only.
• dialog.rs: new DialogInner::end_session_without_ack. When the INVITE or re-INVITE transaction reaches Terminated after we sent a 2xx and no ACK came, it emits Terminated(TerminatedReason::Timeout) and then sends a BYE (best effort; a failure is logged). It is called from handle_invite and handle_reinvite in both InviteDialog and the deprecated ServerInviteDialog.
• For a re-INVITE, whether a 2xx was sent is read right after the final response goes out, because Transaction::cleanup takes last_response on termination.

Nothing changes when the ACK arrives, when the INVITE is rejected or CANCELled, or when the endpoint shuts down (in that case the transaction is not Terminated).

Relation to #127 / #128

This addresses the defects reported in #127 that still reproduce on current main (0.6.12): an ACK that arrives after T4 is swallowed and the dialog stays in WaitAck, and a 2xx that is never ACKed is retransmitted for only about T4 while the session is never ended.

It deliberately keeps a server 2xx in Completed. In rsipstack the transaction layer is the only place that retransmits the 2xx (the dialog layer's accept() sends it once), so keeping it there gives the on-the-wire behaviour RFC 6026 §7.1 asks for — the 2xx is retransmitted until the ACK arrives or 64·T1 passes, and the ACK reaches the TU — without moving retransmission out of the transaction. It does not add the RFC 6026 Accepted state; it is a smaller, targeted alternative to #128 for the observable problem.

Compatibility / risk

• New public field EndpointOption::t2 (default 4 s, the RFC 3261 value), added the same way as the existing t1 / t4 / t1x64 fields. Code that builds EndpointOption without ..Default::default() needs the new field; every construction in the repo already uses the default.
• A UAS dialog whose 2xx is never ACKed now ends with Terminated(TerminatedReason::Timeout) plus a BYE after 64·T1, instead of staying up with no ACK.
• A server INVITE transaction now lives up to 64·T1 (Timer D) after a final response instead of T4; nothing changes once the ACK arrives.
• Over TCP/TLS/WS a 2xx is now retransmitted (T1 doubling to T2) until the ACK arrives, as on UDP.

Checks (on main @ 3ea35cd)

• Red: with the fix reverted and the new tests kept, the no-ACK, re-INVITE and TCP tests fail.
• Green: cargo test → 337 lib + 65 doc passed, 0 failed. rustfmt --check clean on the changed files.