Learning to Simulate Natural Language Feedback for Interactive Semantic Parsing