Weight Squeezing: Reparameterization for Compression and Fast Inference