perf(deploy): keep model resident + pre-warm to eliminate cold reload

diag(deploy): surface GPU visibility + stream Ollama logs to stdout

port change

HF Space deploy — initial