Fix flaky test_dont_steal_long_running_tasks

[crusaderky]

https://github.com/dask/distributed/runs/6159504580?check_suite_focus=true

        actual_total_occupancy = 0
        for worker, ws in self.workers.items():
>           assert abs(sum(ws.processing.values()) - ws.occupancy) < 1e-8
E           AssertionError

long-running tasks are in WorkerState.processing, but don't contribute to WorkerState.occupancy

See

distributed/distributed/scheduler.py

Lines 1955 to 1962 in 198522b

	old = ws.processing.get(ts, 0)
	ws.processing[ts] = total_duration
	if ts not in ws.long_running:
	self.total_occupancy += total_duration - old
	ws.occupancy += total_duration - old
	return total_duration

and

distributed/distributed/scheduler.py

Lines 4781 to 4790 in 198522b

	occ: float = ws.processing[ts]
	ws.occupancy -= occ
	self.total_occupancy -= occ
	# Cannot remove from processing since we're using this for things like
	# idleness detection. Idle workers are typically targeted for
	# downscaling but we should not downscale workers with long running
	# tasks
	ws.processing[ts] = 0
	ws.long_running.add(ts)
	self.check_idle_saturated(ws)

[github-actions]

Unit Test Results

       16 files  ±0         16 suites  ±0   7h 31m 31s ⏱️ + 13m 21s
  2 731 tests ±0    2 648 ✔️  - 1       82 💤 ±0  1 ❌ +1 
21 733 runs  ±0  20 684 ✔️  - 3  1 048 💤 +2  1 ❌ +1 

For more details on these failures, see this check.

Results for commit 68ca3bb. ± Comparison against base commit 198522b.

[fjetter]

Failing test is about a leaking thread, see also #5275

Appears to happen very rarely