Noisy Exemplars Make Large Language Models More Robust: A Domain-Agnostic Behavioral Analysis