Feel Risky to Train Your Language Model on Restricted Data? Meet SILO: A New Language Model that Manages Risk-Performance Tradeoffs During Inference