Talk: Automatic Granularity Control for Efficient Parallelism