Share lazy job scheduling across dataloaders

[ydah]

This refactors lazy job scheduling shared by Dataloader and AsyncDataloader.

• Move selection and enqueueing of the next lazy depth into the base Dataloader
• Use a block for implementation-specific Fiber/Task execution
• Keep Async queue lifecycle management isolated in run_lazy_jobs
• Remove duplicated code and stale TODOs

No behavior change is intended.

[rmosolgo]

Looks good to me!

In general I have become wary of passing blocks in GraphQL-Ruby's runtime methods, because of the performance overhead they add (like the memory footprint of the local binding). But if this turns out to be problem, we could probably replace the yield call with a method call like self.after_lazies(...), and then implement that method in one way or another in the two dataloader classes.