LLM costs need to be attributed per team or feature, but tracing tokens through shared services is hard. What billing setup actually works?