_scheduler: Refactor of queues and resources.

This branch makes the following changes:

  * jobs/job.py: No longer stores any interested resource list

    Jobs are ephemeral again, they only ever exist while they
    are running.

  * queues/queue.py: Revert to only handling lists of elements

    Elements pass through the queues, Queue.harvest_jobs()
    replaces Queue.pop_ready_jobs() and now the Queue stops
    creating jobs as soon as there are not enough resources
    for the job.

    Also removed unused `prepare()` abstract method.

  * jobs/{cachesize,cleanup}job.py: Expose uniform complete callback

    Allows the scheduler to manage resource deallocation for these
    two job completions as a custom thing, at the same phase
    that the Queues take care of their own resource deallocation.

  * resources.py: No longer has knowledge of the job

    Since jobs are ephemeral, they are not a suitable place
    to store the resource identifiers, these must be provided
    by the callers wherever needed.

    Now the main Resources object is owned by the Scheduler
    but shared with Queues, each take care of managing the
    resources of the jobs they create through the same
    resource API.

  * scheduler.py: Reverted to only creating jobs on demand

    This changes the flow of the scheduler such that whenever
    jobs complete, the queues are interrogated for as many
    jobs which can run at the moment but not more; and this
    completely removes the waiting list.

    For the internal cache management jobs, we handle this
    with a little state instead of having a waiting list
    and only launch when the resources permit it.

By abolishing the scheduler waiting list and creating jobs
on demand, we fix the order of element processing and consequently
fix issue #712.