Add routed Modal multi-model deployment

docs: clarify judge inference setup

feat: add embedded llama.cpp and Modal LLM runtimes