From Capabilities to Performance: Evaluating Key Functional Properties of LLM Architectures in Penetration Testing