Forecasting Potential Misuses of Language Models for Disinformation Campaigns--and How to Reduce Risk