Modularize beacon processor scheduler

[eserilev]

Issue Addressed

Partially #6291

Proposed Changes

This PR aims to "modularize" the beacon processor scheduler. The goal is to set up the scheduler in such a way that new scheduling algorithms can be implemented/swapped in.

For now we've decided on exposing an interface that implements the following trait

pub trait Scheduler<E: EthSpec, S: SlotClock> {
    fn new(
        beacon_processor: BeaconProcessor<E>,
        beacon_state: &BeaconState<E>,
        event_rx: mpsc::Receiver<WorkEvent<E>>,
        spec: &ChainSpec,
    ) -> Result<Box<Self>, String>;

    fn run(
        self,
        work_journal_tx: Option<mpsc::Sender<&'static str>>,
        slot_clock: S,
        maximum_gossip_clock_disparity: Duration,
    ) -> Result<(), String>;
}

We then define a SchedulerType enum

pub enum SchedulerType<E: EthSpec, S: SlotClock> {
    PriorityScheduler(priority_scheduler::Scheduler<E, S>),
}

Each variant is a beacon processor scheduling implementation with its own unique scheduling algorithm. Swapping between the variants can be done either at the config level, or with a small code change.

Additional Info

• The concept of work queues is currently specific to the PriorityScheduler variant. Other scheduling algorithms might not use all these different queues to schedule the different WorkEvents and so these queues aren't exposed outside of the PriorityScheduler impl.
• The reprocess event channel is no longer externally exposed. All work events are now sent through the single BeaconProcessorSend channel. I've introduced a new Work::Reprocess type which we then use to schedule jobs for reprocess. The reprocess logic itself is very specific to the PriorityScheduler so it seems to make more sense to give other implementations maximum flexibility.
• There may be some ways to generalize this even further so that different scheduling implementations aren't duplicating similar code/logic... I think we'll find out more once we start actually implementing a new scheduling algorithm