Less is More: Extreme Gradient Boost Rank-1 Adaption for Efficient Finetuning of LLMs