Outlier-Aware Training for Low-Bit Quantization of Structural Re-Parameterized Networks