fix: CANCEL a dropped INVITE after its first provisional response

[tgeorge06]

Fixes #166.

Problem

When the do_invite future is dropped while the dialog is Calling (no response yet), the guard drops the INVITE transaction and sends nothing. A later 180 gets no CANCEL, a later 2xx is neither ACKed nor BYE'd, and a later non-2xx final is not ACKed. The callee keeps ringing, or answers into dead air.

Spec

• RFC 3261 §9.1: "If no provisional response has been received, the CANCEL request MUST NOT be sent; rather, the client MUST wait for the arrival of a provisional response before sending the request."
• RFC 3261 §13.2.2.4: "The UAC core MUST generate an ACK request for each 2xx received from the transaction layer."
• RFC 3261 §15, RFC 5407 §3.1.2: a 2xx to an abandoned INVITE is ACKed and the session it established is ended with a BYE (as fix: ACK and BYE a 2xx that crosses the CANCEL of a dropped INVITE (rebase of #159) #162 does for the Trying / Early drop).

Fix (src/dialog/invitation.rs, src/dialog/invite_dialog.rs)

• DialogGuardForUnconfirmed::drop: Calling now shares the Trying / Early arm. For Calling it still reports Terminated(UacCancel) at once, before spawning, as before.
• The spawned task, for Calling only, first stops the INVITE retransmissions and waits up to 64*T1 for the first response (wait_response, which is wait_final_response with a final_only flag):
  • a provisional: it continues into the existing Trying / Early code unchanged (CANCEL, 2 s settle window, wait for the final response, ACK and BYE a crossing 2xx). Its 64*T1 final-response window starts after the provisional. The second Terminated(UacCancel) that path reports is skipped for Calling, which already reported it.
  • a 2xx: the transaction ACKs it, and the session is ended with a BYE.
  • any other final response, including Timer B's 408: the transaction ACKs it if it came from the peer; nothing else is sent.
  • no response within 64*T1: nothing is sent.
• That path calls InviteDialog::send_cancel, split out of cancel() without its state check: the dialog is already Terminated when a Calling drop sends its CANCEL. For Trying / Early the guard already matched the state, so the CANCEL sent is the same as before.
• The 2xx-then-BYE tail moves into bye_if_2xx, shared by both paths.

Contract / coverage

• Abandoning a call before any response reports Terminated(UacCancel) at once, as before, and once. Nothing is sent before a provisional, and the INVITE is no longer retransmitted.
• The first response then gets: a CANCEL for any 1xx (100 included, as in Trying), an ACK and a BYE for a 2xx, the transaction's ACK for 3xx-6xx. No Confirmed is reported for an abandoned call.
• Same code on UDP, TCP, TLS and WS; on reliable transports there were no retransmissions to stop. No tokio-only code: the task uses the crate::platform primitives already used by this arm.
• Unchanged: the Trying / Early drop (test_2xx_before_cancel_response_is_acked_and_byed, test_2xx_after_cancel_response_is_acked_and_byed, test_2xx_after_cancel_settle_window_is_acked_and_byed, test_cancel_answered_487_sends_no_bye pass unchanged), the Confirmed drop (hangup), InviteDialog::cancel() for applications, and do_invite_async, which does not use the guard.
• Not changed, the same as in the existing Trying / Early path: only the first 2xx of a forked INVITE is BYE'd; with EndpointOption::auto_ack_2xx = false the 2xx is not ACKed before the BYE; a reliable provisional (100rel) gets no PRACK before the CANCEL; the task does not watch the endpoint's cancellation token. These belong to both paths and are left for separate changes.

Tests

src/dialog/tests/test_cancel_2xx_race.rs, with the file's raw UDP peer (default T1 = 500 ms). run_dropped_before_provisional(first) drops do_invite right after the INVITE arrives, asserts Terminated(UacCancel) within 200 ms and that the peer receives nothing for 1.1 s (over twice T1), then sends first:

• test_dropped_before_provisional_is_cancelled_after_the_180: a CANCEL with the INVITE's CSeq follows, then 200 to it and 487 to the INVITE, whose ACK follows; no BYE.
• test_dropped_before_provisional_2xx_is_acked_and_byed: the 200 is ACKed and a BYE follows, both in the 2xx's dialog.
• test_dropped_before_provisional_final_failure_is_acked_only: the 486 is ACKed; no BYE.

Each also asserts that no second Terminated and no Confirmed is reported. On main:

test_dropped_before_provisional_is_cancelled_after_the_180 ... FAILED
  panicked at src/dialog/tests/test_cancel_2xx_race.rs:34:33: timeout waiting for CANCEL
test_dropped_before_provisional_2xx_is_acked_and_byed ... FAILED
  panicked at src/dialog/tests/test_cancel_2xx_race.rs:34:33: timeout waiting for ACK
test_dropped_before_provisional_final_failure_is_acked_only ... FAILED
  panicked at src/dialog/tests/test_cancel_2xx_race.rs:34:33: timeout waiting for ACK

A 2xx that crosses the deferred CANCEL runs the same code as run_crossing_2xx, so it is not tested again here.

Checks

• cargo test --features bench: 369 lib tests passed, 0 failed (366 on main), and 65 doc tests passed. Plain cargo test builds too.
• cargo check --no-default-features --features platform-embassy: no warnings, as on main.
• cargo clippy --features bench --all-targets: the same 114 warnings as on main, none in the changed code.
• rustfmt --check on the changed files: clean.
• Merges cleanly with fix: keep confirmed dialogs confirmed on 1xx to in-dialog requests #147, fix: notify dialog state only for transitions that are applied #148 and the late-ACK routing PR (fix: route the ACK of a 2xx to the INVITE with the same CSeq #167); with all of them merged all tests pass.

[shenjinti]

Adopted via #171 (clean rebase, no conflicts). Merged with all tests green — thank you!