S$^{2}$FT: Efficient, Scalable and Generalizable LLM Fine-tuning by Structured Sparsity