WIP: ass_cache: experiment with non-blocking optimizations for the LR…

…U list

- Instead of adding/promoting items in the LRU list within ass_cache_get, add them to a simpler singly-linked lock-free list of items used this frame
- In ass_cache_cut, add/promote the contents of the list from the previous frame
- Take locks as little as possible (by using double-checked locking patterns, making more things atomic, etc)
- Shard locks by bucket (this should really change to be by thread, via a mechanism suggested by MrSmile)