Rejuvenating Low-Frequency Words: Making the Most of Parallel Data in Non-Autoregressive Translation