Add spec for better loads: dataloading, prototype new loads code

[rmosolgo]

Implementing loads: has had a smell in it:

graphql-ruby/lib/graphql/schema/argument.rb

Line 338 in c4f07cd

context.dataloader.run_isolated do

graphql-ruby/lib/graphql/schema/argument.rb

Line 359 in c4f07cd

context.dataloader.run_isolated do

It's there because some code requires the argument values to be returned right away -- it doesn't use Dataloader (properly?).

This smell manifests the problem described in #5462 -- I added a spec for that issue in cf3a3f8

This problem also surfaced in #5479 because, when using Async, run_isolated isn't actually isolated: Async uses a top-level pool, so a pause in one place in the code can resume code elsewhere.

The smell was also called out in #5389 (comment).

When working on #5389, I didn't actually rework this code. But I think it's actually a good place to try out that new flow in an incremental way.

Fixes #5462

TODO:

• Remove another Argument-related run_isolated call
• What about the mutation-related run_isolated call?
• Find a better solution than while ...; dataloader.yield; end -- some proper way for expressing or communicating dependencies in the code
• Consider rooting out after_lazy usage in loading and authorization
• Benchmark