Text this: Tackling the Data-Parallel Load Balancing Bottleneck in LLM Serving: Practical Online Routing at Scale