G-DIG: Towards Gradient-based DIverse and hiGh-quality Instruction Data Selection for Machine Translation