Uncovering the Strategic Reasoning Limitations of LLMs via Game-Theoretic Evaluations