LLM as Effective Streaming Processor: Bridging Streaming-Batch Mismatches with Group Position Encoding