remove potential false sharing on counters

[bencsikandrei]

add a simple counter that aligns on 64B (size of a chache line) use that instead of having 8 B uint64s consecutive and touched by different threads the impact is larger if the processing itself is fast, otherwise the bottleneck is somewhere else, still, not invalidating cache lines is nice

[stefankarschti]

Great stuff! The only drawback is the extra memory usage, but that can be avoided by grouping elements in structs aligned to 64B.