How to Compare Quantized LLMs Fairly in llama.cpp: Same Bytes, Same Windows, Same GPU