Refactor Scheduler to class

[mrocklin]

Previously the scheduler was a set of coroutines and queues that should be run together. This was successfully done in the Executor class but rarely run separately.

Now we package up the coroutines, queues, and various state collections (dask, dependents, etc.) into a single Scheduler class. We do this with an eye to run a scheduler on a separate machine from where the executor runs. This is the first step in a few large features, notably running a client executor on a local laptop while having the scheduler on the cluster and also having multiple executors connected to the same scheduler.

The Scheduler and Executor connect by two queues (incoming and outgoing). At the moment these queues are connected directly but in the future we might interpose a socket stream.