Data Valuation for LLM Fine-Tuning: Efficient Shapley Value Approximation via Language Model Arithmetic