Counting Without Running: Evaluating LLMs' Reasoning About Code Complexity