Fine-tuning LLMs to 1.58bit: extreme quantization made easy