Building a Subspace of Policies for Scalable Continual Learning