Quantization, Explained: From FP32 to INT4, and What Each Precision Costs You