OWQ: Lessons learned from activation outliers for weight quantization in large language models