JoinHandle destructor should not panic when dropping output

[Darksonn]

In Tokio, panics are generally caught and not propagated to the user when dropping the JoinHandle, however when dropping the JoinHandle of a task that has already completed, that panic can propagate to the user who dropped the JoinHandle. That happens here:

tokio/tokio/src/runtime/task/harness.rs

Lines 167 to 193 in 4eed411

	pub(super) fn drop_join_handle_slow(self) {
	let mut maybe_panic = None;
	// Try to unset `JOIN_INTEREST`. This must be done as a first step in
	// case the task concurrently completed.
	if self.header().state.unset_join_interested().is_err() {
	// It is our responsibility to drop the output. This is critical as
	// the task output may not be `Send` and as such must remain with
	// the scheduler or `JoinHandle`. i.e. if the output remains in the
	// task structure until the task is deallocated, it may be dropped
	// by a Waker on any arbitrary thread.
	let panic = panic::catch_unwind(panic::AssertUnwindSafe(|| {
	self.core().stage.drop_future_or_output();
	}));
	if let Err(panic) = panic {
	maybe_panic = Some(panic);
	}
	}
	// Drop the `JoinHandle` reference, possibly deallocating the task
	self.drop_reference();
	if let Some(panic) = maybe_panic {
	panic::resume_unwind(panic);
	}
	}

Note that the unset_join_interested call can only return an error if the task has already completed, so it is the output being dropped — not the future.