fix(voice): bound EndCallTool wait for tool-reply playout

[usmamalik345]

Summary

• _delayed_session_shutdown already wraps the wait for speech_created in asyncio.wait_for(..., timeout=5.0), but then does a bare await speech_handle
• if that handle never completes, finally never runs and ctx.session.shutdown() is skipped — room/delete_room cleanup never happens (SIP can stay up)
• wrap the speech-handle await in the same TOOL_REPLY_TIMEOUT, keep the existing warning + shutdown path

Closes #5096 (timeout asymmetry only; not Gemini instruction-following).

Testing

• pytest tests/test_tools.py::test_delayed_session_shutdown_times_out_when_speech_handle_hangs — fails on main (hangs), passes with the fix
• also covers: speech_created never fires; handle completes normally
• ruff check / ruff format clean on touched files

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[usmamalik345]

Addressed the Devin review note:

• timeout path now interrupt(force=True) on the hanging tool-reply handle (when we have one), then shutdown(drain=False) so drain/close is not blocked on that same unfinished speech
• successful playout still uses default shutdown() (drain=True)
• tests assert drain=False on both timeout paths and drain=True when the handle completes normally