Simple Embodied Language Learning as a Byproduct of Meta-Reinforcement Learning