How LLM Inference Serving Works: KV Cache, PagedAttention and Continuous Batching