From 300KB to 69KB per Token: How LLM Architectures Solve the KV Cache Problem