Multi-Bin Batching for Increasing LLM Inference Throughput