fix: stop quit() from cancelling its own owning task through a gathered tool call

[AmirF194]

The "quit" tool's coroutine runs as a child task of ToolExecutor.execute_async's own asyncio.gather() (bub/tools.py), so asyncio.current_task() inside it is that child task, not the _run_message task registered in controller.active_tasks. quit()'s self-skip check compares against the wrong task, cancels its own ancestor, and awaits it, but that ancestor is itself awaiting the same gather() this code is a child of, so CPython recurses cancelling it without bound.

Fixed by tracking the owning task in a contextvar set once at the top of _run_message. Child tasks, including the ones execute_async's gather spawns, inherit a copy of that context, so quit() resolves the real owning task regardless of how many awaits deep it's invoked from.

Reproduced against the real ChannelManager/ToolExecutor classes: same RecursionError and wedged event loop on main (990 recursive _GatheringFuture.cancel() frames), clean exit with 0 tasks cancelled on this branch, both confirmed on Python 3.12 and 3.14. The regression test runs the scenario in a subprocess with an OS-level timeout, since the hang also defeats asyncio.wait_for's own cancellation and can't be bounded from inside the event loop itself. Full suite, ruff, and mypy pass (mypy has one pre-existing unrelated error in settings.py that's on main too).

[lodyai]

The ContextVar approach looks like a good minimal fix: the task to exclude is the owning turn, which may differ from asyncio.current_task() inside a gathered tool call.

A few suggestions:

Scope _owning_task with a token and try/finally: token = _owning_task.set(asyncio.current_task()), then _owning_task.reset(token) in finally. This is not necessary for the current fresh-task entrypoint, but keeps the context scoped correctly if _run_message is ever directly awaited or nested.

[AmirF194]

Good catch, pushed the token/finally scoping so the contextvar clears itself after each turn instead of persisting past it.