Large language models for risk-of-bias assessment in randomised clinical trials—a comparative validation study