Add Response.thoughts()

I see that you have an open issue for visible thinking tokens (#770), and that you had to abandon your first attempt. In the meantime, would it be possible to add a thoughts() or thinking() method to the Response class that returns the thinking tokens?

I'm aware that I can dig the thinking tokens out of the json response, but I'd like a less brittle, model-agnostic way of doing it. (In case it's not clear, I'm using llm as a python library rather than as a cli tool, and I'm trying to get at Anthropic's thinking tokens.)

Maybe this is too crude, but one possibility would be to add self.thoughts = None to _BaseResponse, and dig it out of the json object in ClaudeMessages and AsyncClaudeMessages:

class ClaudeMessages(_Shared, llm.KeyModel):
    def execute(self, prompt, stream, response, conversation, key):
        (...)
        if stream:
                (...)
                # This records usage and other data:
                response.response_json = stream.get_final_message().model_dump()
                if <thinking>:
                    response.thoughts = response.response_json[(...)]
        else:
            (...)
            yield prefill_text + text
            response.response_json = completion.model_dump()
                if <thinking>:
                    response.thoughts = response.response_json[(...)]
        self.set_usage(response)


class AsyncClaudeMessages(_Shared, llm.AsyncKeyModel):
    async def execute(self, prompt, stream, response, conversation, key):
        (..)
        if stream:
            (...)
            response.response_json = (await stream_obj.get_final_message()).model_dump()
            if <thinking>:
                response.thoughts = response.response_json[(...)]
        else:
            (...)
            yield prefill_text + text
            response.response_json = completion.model_dump()
            if <thinking>:
                response.thoughts = response.response_json[(...)]
        self.set_usage(response)

I can put up a PR if you think this is worth pursuing.

Start by reading the Response class and _BaseResponse, then trace ClaudeMessages and AsyncClaudeMessages to see how response_json is populated for streaming and non-streaming completions. Define how a model-agnostic thoughts() or thinking() method should expose Anthropic thinking tokens in both paths, and verify that existing response behavior remains intact.