Exploring the Effectiveness of Multi-stage Fine-tuning for Cross-encoder Re-rankers