Scaling Up Membership Inference: When and How Attacks Succeed on Large Language Models