Isn't it quite unfriendly to make the caller implement the main loop, and require periodic polling of the scheduler in order to run jobs? When I have implemented things like this before, I kept tuples of (time_of_next_run, job) in a heap. Then the scheduler had a loop like:

  while True:
      now = time.time()
      while job_heap[0][0] < now:
          time, job = heapq.heappop(job_heap)
          job.run()
          heapq.heappush(job_heap, (job.next_time(), job))
      time.sleep(job_heap[0][0] - time.time())

This reduces the amount of polling needed, since you sleep until the next job is ready.