Best Practices for Distilling Large Language Models into BERT for Web Search Ranking