PolyTask: Learning Unified Policies through Behavior Distillation