Training-Free versus Training-Based Intent Classification in LLMs: Accuracy, Robustness, and Failure Modes