Use concurrent_task_barrier() instead of synchronization

[seberg]

Especially, the kmeans code has communication synchronization calls (but also knn, and others).

As @magnatelee mentioned, it should now be possible to replace these with TaskContext::concurrent_task_barrier() from legate (I am not sure if the handle.get_comms().barrier() would also do something, but I expect not).

This solution would be much preferred; on the implementation side, I do think it requires threading through the context (although, I am not sure the API needs anything from the TaskContext really).