DenseSet::Grow suffers from slow memory latency and it has a good potential for optimization.

introduce resharding algorithm that moves items to new buckets after the resize, but it does it in chunks of kBatchSize.

Similar to #3863

1. Must add benchmarking code to string_set_test to evaluate the performance
2. The benchmarking code must only evaluate Grow - which is not simple thing to do. For that it must add, say 2^15 items without mesuring and then use additional Add to trigger Grow event (when we cross 2^k we grow). The additional Add must be under measurement. See BM_AddMany for example, how we control timing with PauseTiming/ResumeTiming.
3. Once the benchmark exists you can run it with ./string_set_test --bench --benchmark_filter=.*BM_Grow - should be done in opt mode.
4. Now it is possible to improve the implementation of grow - introduce a state machine that takes a batch of buckets and iteratively reshards all the elements in them (GrowBatch). Grow should iterate over all old buckets and call GrowBatch.

I would expect ~50% CPU reduction for large sets