Practical Considerations for Fine-Tuning BERT-Based Language Models in Health Research: Lessons from Classifying Anti-Vaccine Posts on Social Media