Accelerating Distributed Deep Learning Training with Compression Assisted Allgather and Reduce-Scatter Communication