Zero-Waste Agentic RAG: Designing Caching Architectures to Minimize Latency and LLM Costs at Scale