Cognitive Decision Routing in Large Language Models: When to Think Fast, When to Think Slow