Training with Quantization Noise for Extreme Model Compression