On the Emergence of Whole-body Strategies from Humanoid Robot Push-recovery Learning