SARIA: Persistence-Aware KV Caching with Coordinated DRAM--SSD Reload for LLM Agents