This avoids the potential to race with the triggering coming from
task_cancel, because we first set the cancelled flag, and only THEN
take the lock and iterate over the inserted records. Because of this we
could: T1 flip the cancelled bit; T2 observes that, and triggers
"immediately" during installing the handler record. T1 then proceeds to
lock records and trigger it again, causing a double trigger of the
cancellation handler.
resolves https://github.com/swiftlang/swift/issues/80161
resolves rdar://147493150
