Cleanup/refactor the scheduler once again, and write up design doc

[MingweiSamuel]

Scheduler is fixed for now, but a bit of a mess and overcomplicated #756

Once I have time I should try to clean it up and write up the design in a doc

Partial list of scheduler-related issues:

• anything about replay
• Ops for scheduling ticks and for deferring dataflow across ticks #623 feat: add defer_tick_lazy() which is the same as defer_tick() except that it is lazy #945
• make edge references cause scheduling #1175 (patch fix by inserting persist()s)
• Singletons may reference 'tick stateful operator state that wasn't reset for the tick (due to the subgraph not being scheduled), resulting in incorrect behavior #1298

[MingweiSamuel]

#97

[MingweiSamuel]

Now #1500 for flo semantics