Not All Bias is Bad: Balancing Rational Deviations and Cognitive Biases in Large Language Model Reasoning