Sync workers with global lock

[kevinzou1982]

compare to my last pull request 1, use global sync lock instand of lock_counter and release_counter in tensorflow. speed up counter access and prevent not zero problem. 2, fix a sync bug. Clipped_ppo_agent for example if int(batch.size / self.ap.network_wrappers['main'].batch_size) is different between each workers, some worker will be in acting phase and other workers will be block in wait_for_all_workers_barrier