Exploring and Improving the Spatial Reasoning Abilities of Large Language Models