MultiPL-E: A Scalable and Extensible Approach to Benchmarking Neural Code Generation