Unify execution metadata objects

[rmosolgo]

For each key in the response, we create one Context and one-or-zero FieldResult/SelectionResult. The leads to a lot of extra GC work, see this benchmark which highlights graphql overhead:

#861 (comment)

We can remove this overhead by using Context#value/Context#value= to manage results.

Another advantage is that you can backtrack through the results using context.parent.value, which will fix #881.

Also fixes #882.

Finally, this lays the groundwork for streaming execution. Since ctx objects track their underlying objects, they can be used to resume execution at any time. This could reduce TTFB with graphql-client.

TODO:

• Address todos in code, smooth out the ad hoc conditionals which were scattered around to make the tests pass
• simplify query/context.rb:
  • (as much as i could) Merge Query::Context & Query::Context::FieldResolutionContext
  • Move to a new class which describes its current role May as well leave it, break less stuff
  • Move Context.flatten to somewhere that makes sense (inside Execution?)
• Rename variables in execution code to match the kind of objects they are
• Add a test for backtracking context