Instance-level Randomization: Toward More Stable LLM Evaluations