Add config method to parse prose output into structured JSON via a lightweight model

Follow-up to #1007. Add a config method that runs a small/cheap LLM to coerce a cog's prose prompt into a call to the structured JSON method.

This would allow users to set something like: "Output a Json with the names, followed by the emails" and the lightweight model infers the corresponding schema and wires it into the structured-output method from #1007, so the primary model's response is forced to that shape. No hand-written schema required.

config do
  chat(:extract) do
    generate_schema_with "gpt-4o"
  end
end

Start by reading the structured JSON method described in follow-up issue #1007 and the configuration pattern shown in this issue. Trace how a cog's prose prompt reaches the primary model, then determine how generate_schema_with "gpt-4o" should connect schema inference to structured output. Done means users can provide the prose prompt without a hand-written schema and receive the requested JSON shape.