Registration worker can fall behind during initial sync

[ptrus]

Registration worker can fall behind with epoch registrations, mostly noticeable in case where the node exists sync mode before it is actually caught up (related: #2861) - where one epoch registration attempt could take more time than it takes to sync one epoch worth of blocks.

Consider using an overwriting channel (RingChannel) in WatchEpochs cases where it is safe to skip epochs - the registration worker & possibly others.