Exploring Reasoning Biases in Large Language Models Through Syllogism: Insights from the NeuBAROCO Dataset