Improper Learning for Nonstochastic Control