Accelerating Transformer Inference for Translation via Parallel Decoding