Scaling Retrieval-Based Language Models with a Trillion-Token Datastore