Refactor `dacluster` to start controller and engines separately

From @minrk :
"""
You can start the controller and engines separately, which makes restarting engines easier (though they will get new IDs). I typically do it this way:

ipcontroller
ipcluster engines —engines=MPI -n 8

Then I can kill/restart the engine processes.
"""

We should refactor `dacluster` to use this pattern, so that `dacluster restart` simply kills the engine processes and then fires up a new set of them.