Dissecting Batching Effects in GPT Inference