Pipeline Parallelism in SGLang: Scaling to Million-Token Contexts and Beyond