What is the impact of sharing recalculation on data load performance for objects with private OWD, and what techniques (e.g., deferred sharing calculation) mitigate this during large migrations?