Fixes the timing dependence described in the issue: the test cancelled the parent task from outside after await asyncio.sleep(0.01), and lost the cancellation message whenever the loop stalled for more than 10 ms before the body task first ran (the mechanism, and a reproducer that stalls the loop with time.sleep(), are in the issue).

The test now synchronizes on two events instead of sleeping. The child signals from its finally block that the group is cancelling, which by then has cancelled the parent task and is waiting for the child, and waits there until the test releases it. Only then does the test cancel the parent task from outside. The ordering is fixed by the callback queue, not by time, so it holds on any build speed and for both task implementations.

Checked:

• The rewritten test passes for TestTaskGroup and TestEagerTaskTaskGroup, 10 of 10 repeated runs, and the whole test_asyncio.test_taskgroups module passes.
• With the asyncio.TaskGroup swallows cancellations coming from outside of it #155433 fix reverted locally, the rewritten test still fails in both classes (CancelledError not raised), so it keeps guarding the bug it was written for.
• The event-based ordering was also verified in a standalone harness with the same 20 ms loop stall that makes the current test fail.

Since the original test was added by the gh-155433 fix, this likely wants the same backports.

• Issue: test_taskgroup_cancel_keeps_outer_cancellation is timing-dependent and fails on slow (TSan) CI #157157