"Give Me BF16 or Give Me Death"? Accuracy-Performance Trade-Offs in LLM Quantization