Built-in way to have CoT/Thinking tokens accessible from generator output

[swerner]

Been experimenting with R1 models locally with https://ollama.com/MFDoom/deepseek-r1-tool-calling:latest - was able to get the tool calling working pretty reliably with the 32b model. Only problem was that you lost the thinking tokens which could also be useful.

I've also found myself setting up my outputs to have somewhat of a CoT feel, though with outputs. Example here: https://github.com/sublayerapp/sublayer_actions/blob/main/.github/workflows/daily_action_suggestion/generators/action_ideas_generator.rb

llm_output_adapter type: :list_of_named_strings,
    name: "action_ideas",
    description: "List of sublayer action ideas with usefulness scores",
    item_name: "idea",
    attributes: [
      { name: "title", description: "The title of the action idea" },
      { name: "description", description: "A brief description of the action" },
      { name: "usefulness_score", description: "A score from 1-10 indicating the usefulness of the action with 10 being the best" }
    ]

Thinking that with these smaller models matching sonnet 3.5 on benchmarks and being able to easily run on a macbook, along with really good results coming out of gemini-2.0-thinking-experimental we'll want to take advantage of these improvements.

I think I've also come up with a workable solution for https://github.com/sublayerapp/sublayer/issues/92 as well to tie in to this...