Assessing Machine Learning Robustness in MRS Quantification: Impact of Training Strategies on Out-of-Distribution Generalization