Evaluating the Efficacy of LLM-Based Reasoning for Multiobjective HPC Job Scheduling