Add status messaging and endpoint to allow for client-side users to reason about model initialization and life cycle.

[tstescoTT]

Logs during setup, especially tt_metal_cache generation on 1st run, for example:

inference_api_server - ERROR - Model backend is hanging. time_since_response:=844.313969373703, time_since_status_msg:=844.30121231079

Are confusing because the model is timing out from the inference server perspective, but is still initializing so this is expected behavior.

Correct way to deal with this is to use the status queue and push messages for model init process to the inference server, the server can in turn expose a status endpoint for downstream consumers to understand when the model is ready.

[tstescoTT]

Should explore usage in vLLM.