Accelerating Communication in Deep Learning Recommendation Model Training with Dual-Level Adaptive Lossy Compression