Performance regression in Stripe's codebase from Opus::Enum DSL

[jez]

Problem

The changes introduced in #1480 have caused a significant performance regression in typechecking Stripe's codebase. Basically, by turning this:

class MyEnum < Opus::Enum
  A = new
  B = new
end

into this:

class MyEnum < Opus::Enum
  class A$1 < MyEnum; ...; end
  class B$1 < MyEnum; ...; end
end

It makes it so that none of the enum values can be resolved on the first pass of resolver, so they all have to be enqueued as AncestorResolutionItems.

Proposed solution

We believe that if the DSL pass would be rewritten to emit code like this:

class MyEnum < Opus::Enum; ...; end
class MyEnum::A$1 < MyEnum; ...; end
class MyEnum::B$1 < MyEnum; ...; end

Far fewer things would have to be queued up in resolver, which would make the time spent typechecking pay-server smaller.