Limits of PRM-Guided Tree Search for Mathematical Reasoning with LLMs