Improving Generalization in Reinforcement Learning Training Regimes for Social Robot Navigation