Towards Benchmarking and Improving the Temporal Reasoning Capability of Large Language Models