The case for 4-bit precision: k-bit Inference Scaling Laws