Demystifying LLM Serving Infrastructure: How PagedAttention and Continuous Batching Scale Inference