Inside vLLM: Anatomy of a High-Throughput LLM Inference System