Resources are ignored if a task is cancelled and then added again

[crusaderky]

1. The client submits a task without resource restrictions -> the task enters WorkerState.ready

2. The client cancels the task before it can start executing -> the task is forgotten, but its key is still in WorkerState.ready. This is normally ok thanks to
   

   distributed/distributed/worker_state_machine.py

   Lines 1596 to 1604 in a05cc38

   	while self.ready and len(self.executing) < self.nthreads:
   	_, key = heapq.heappop(self.ready)
   	ts = self.tasks.get(key)
   	if ts is None:
   	# It is possible for tasks to be released while still remaining on
   	# `ready`. The scheduler might have re-routed to a new worker and
   	# told this worker to release. If the task has "disappeared", just
   	# continue through the heap.
   	continue

3. The client resubmit the task, with the same key, this time with resource restrictions.
   The task key is now both in WorkerState.ready and WorkerState.constrained.

4. Due to other tasks in the queue, the task reaches the top of WorkerState.ready before it reaches the top of WorkerState.constrained.

5. The task starts executing, completely bypassing resource restrictions. WorkerState.available_resources may become negative for as long as the task is running.

Proposed design

Reimplement ready and constrained as HeapSets. This also solves #6137.

[crusaderky]

I've commented out a wealth of assertions in worker_state_machine.py because of this. Search for this ticket number.