Mitigating Outlier Activations in Low-Precision Fine-Tuning of Language Models