LLM API Cost Structure for Agent Fleets: Empirical Benchmarks for Caching, Routing, and Batching at 100+ Concurrent Tasks