Cognitive Biases in Large Language Models for News Recommendation