Developing a Recommendation Benchmark for MLPerf Training and Inference