What strategies did you implement to optimize PySpark transformations and reduce ETL runtime by 25% at Featherlite Office Systems?