Towards Identifying Social Bias in Dialog Systems: Framework, Dataset, and Benchmark