Eliminate dynamic dispatch from Language::Visitor and Analysis::AST::Visitor

[rmosolgo]

In analysis, Kernel#public_send was a big delay (and it was own time -- overhead -- not the methods that end up getting called). It could mostly be eliminated by some heavy-handed codegen. It seems to me that it pays off:

Before (3.6% own time)

stackprof before-stackprof.dump --method=Kernel#public_send
Kernel#public_send (<cfunc>:1)
  samples:  3518 self (3.3%)  /   52917 total (49.9%)
  callers:
    232534  (  439.4%)  GraphQL::Language::Visitor#visit_node
    22700  (   42.9%)  GraphQL::Analysis::AST::Visitor#call_analyzers
    4106  (    7.8%)  GraphQL::Schema::Warden.visible_entry?

After (0.5% own time)

stackprof last-stackprof.dump --method=Kernel#public_send
Kernel#public_send (<cfunc>:1)
  samples:   507 self (0.5%)  /   44524 total (47.5%)
  callers:
    44524  (  100.0%)  GraphQL::Language::Visitor#visit
    3827  (    8.6%)  GraphQL::Schema::Warden.visible_entry?

TODO:

• Clean up the new Language::Visitor code. I think those visit_{...}_children methods could be dynamically created unless the method was already manually defined, and that would remove the need for the custom method name argument. Also reconsider the method body organization to make sure stuff that can be de-duplicated is.
• Generate methods for Analysis::AST::Visitor to reduce duplication?
• Consider an OO approach to this, for example, an API like node.visit # visit this node and all children recursively. It may be just a matter of ergonomics 🤷