Evaluation Pitfalls and Sparsity Limitations in LLM-based Confidence Estimates for Classification