Fairness Testing in Retrieval-Augmented Generation: How Small Perturbations Reveal Bias in Small Language Models