Evaluating and Enhancing Large Language Models for Conversational Reasoning on Knowledge Graphs