Mind the Memory Gap: Unveiling GPU Bottlenecks in Large-Batch LLM Inference