Match the ollamarunner and OpenAI semantics: raw, full-vocab log-softmax with the top-K ranked by probability. Skipped on the GPU when the request doesn't ask for logprobs so decode doesn't pay for it otherwise.
