Mitigate value type torn writes during LRU update

[bitfaster]

trafficstars

• Since update and delete are performed within a lock, threads pushing items through the queues will not see the torn write.
• Reads during update are susceptible. Key concern here is to avoid a lock during the read because this will lead to lock contention for high frequency reads.
• TODO: the new item must be enqueued:
  • If it is, the queue is polluted wasting some capacity (potentially evicting valid items) and the old item will then age out.
  • If not, wasAccessed in the queue will not reflect the dictionary and the policy is broken
  • Alternatives here would be
    • To add the new item to the dictionary (so readers see the new value without tearing), then update the old item, the update the dictionary a second time.
    • Somehow make the value in the queue updatable