Build and test a token-aware, fair limiter

Three pieces sit on top of each other: a fake model that is slow like a model and rate-limited like a provider, the limiter that decides when a call may go, and the queue that decides who goes next. The decision that matters most is running all of it on a virtual clock: real async code with real asyncio.sleep, Lock and Semaphore, but sleeps finish instantly in time order. Without it each experiment would take ten real minutes per configuration and give a different answer every run.