When Numbers Tell Half the Story: Human-Metric Alignment in Topic Model Evaluation