Mitigate value type torn writes during LRU update (enqueue)

[bitfaster]

During update check if the value is write atomic

• If it is, just update it in place
• Else make a new item, update the dictionary (now the item can be read after without tearing), mark the old item as removed and enqueue the new item
  • Queue cannot be updated in place, so we need to pollute with a dupe for this key in this case

The disadvantage of this approach is that for a large number of updates, the effective size of the cache will be reduced. This is not intuitive, and for very heavy update workloads would likely lead to worse write performance than locking on read/lock contention.

[coveralls]

coverage: 99.141% (+0.01%) from 99.13%
when pulling 9f751d0 on users/alexpeck/tornwrite
into 5483cbf on main.

[bitfaster]

Superseded by #593