PhysicalCoreID() is a major bottleneck

In some use case, `PhysicalCoreID()` called from `recordTick()` is responsible for about 30% of the total CPU consumption of the application, and for even larger share of RocksDB's CPU consumption.

In our use case, just reverting #2258 helped. We don't call RocksDB from too many threads.

It would be nice to have an option to choose either process-local or thread-local stats.

Another approach that could satisfy all use cases is to stripe `StatisticsData` into NCPU * 4 (for example) parts, and update a randomly chosen one in `recordTick()`.