Provide lock free version of write cache

[laa]

Currently we use lock manager to prevent consistency corruption of write cache, but write groups are pretty small and we can use transactional memory approaches based on CAS to prevent contention between threads and improve CPU cache reuse.

From previous experience we usually get 1.5-2 times performance speed up in such cases.

[laa]

We got 4 times speed up on 8 threads on 8 cores. It is good for beginning but later we should improve scalabilty factor to 0.7 IMHO.