BroadcastScheduler

We have landed the prototype BroadcastScheduler for improved scaling of DirectView-style message patterns (same request to all engines), but there's still quite a bit to do. Still to do to bring this to maturity:

1. document differences between broadcast and direct view, tuning parameters, etc.
2. ensure broadcast view passes all direct view tests (some noted in skips in test_view_broacast)
3. include benchmarks comparing dview/broadcast
4. consider 'automatic' coalescing behavior rather than explicit
4. when bcast reaches maturity, consider *replacing* directview with broadcast, as it should be better ~across the board, and only worse when everything is super fast (small messages, small number of engines)