Use AtomicInteger instead of unsafe increment on volatile fields

[stsypanov]

I've found several cases where we increment numeric field declared volatile. Such explicit increment may bring incorrect behaviour, as increment operation itself is not atomic.

[jhoeller]

Reviewing the three affected places, it turns out that all of those are effectively evaluation counts with no need for absolute correctness attached. As far as I can tell, all of those checks would still work correctly even if an occasional concurrent count update has been lost.

That said, this is nevertheless worth addressing, so I'll merge this into master. There is just no strong need for a backport, so I'll rather keep the existing minimal locking in place and leave the potential runtime side effects of stronger locking up to 5.2 only.