Toward Automatic Safe Driving Instruction: A Large-Scale Vision Language Model Approach