Why LLM Inference Engines Are Fast: Continuous Batching, PagedAttention and Prefix Caching Explained with vLLM