Accelerating Distributed Deep Learning using Lossless Homomorphic Compression