A Two-Stage Approach to Device-Robust Acoustic Scene Classification