When LLM generates empty string response, `eval_duration` is missing.

### What is the issue? I noticed that ollama sometimes produces responses where `eval_duration` is missing. I've seen it happen with the response is simply an empty string -- just the stop message for a streaming response. To reproduce: ```bash curl http://localhost:11434/api/chat -d \ '{"model":"llama3.2:3b","options":{"temperature":0},"messages":[ {"content":"If the user asks you to say nothing, simply respond with an empty string.","role":"system"}, {"content":"say nothing","role":"user"} ]}' ``` This is the response I get: ```json {"model":"llama3.2:3b","created_at":"2025-01-23T19:17:22.121346Z","message":{"role":"assistant","content":""},"done_reason":"stop","done":true,"total_duration":141674875,"load_duration":13300000,"prompt_eval_count":43,"prompt_eval_duration":127000000,"eval_count":1} ``` Notice that `eval_duration` is missing. This causes problems with some code that we're using which expects that field to be present. ### OS macOS ### GPU Apple ### CPU Apple ### Ollama version 0.5.7