Tempo: Accelerating Transformer-Based Model Training through Memory Footprint Reduction