PRECOG: PREdiction Conditioned On Goals in Visual Multi-Agent Settings