Two of these caches skip the LLM call for $0; the third makes the call cheaper. Here is what prompt caching, semantic caching, and gateway exact match each catch, and how to layer all three.