scheduler: move OpenAI recording after runner load/release

[doringeman]

If the runner is defunct, then loader.load will evict it, which means it'll also remove the recorded OpenAI requests/responses for that model. Hence, record the request after loader.load returns so it doesn't get removed in the loading process.

Without this patch, following these steps will result in a lost pair of request-response, the first one after the runner exited/was killed and had to be evicted because it was defunct.

$ MODEL_RUNNER_PORT=8080 make run # in a separate terminal

$ MODEL_RUNNER_HOST=http://localhost:8080 docker model run ai/smollm2 hi

$ curl -s http://localhost:8080/engines/requests\?model\=ai/smollm2 # all good, 1 record

$ pkill com.docker.llama-server

$ MODEL_RUNNER_HOST=http://localhost:8080 docker model run ai/smollm2 hi

$ # See `ERRO[0011] Model ai/smollm2 not found in records - 200` in the terminal where model-runner is running.
# This is because the request was recorded, then `loader.load` removed it, 
and then via a deferred function we attempted to record the response, but no matching request was found.
# Hence, we lost the first pair of request-response.

$ curl -s http://localhost:8080/engines/requests\?model\=ai/smollm2
No records found for model 'ai/smollm2'

[p1-0tr]

LGTM