ERASER: A Benchmark to Evaluate Rationalized NLP Models