Toward Best Practices for Training Multilingual Dense Retrieval Models