From scratch to silver: Creating trustworthy training data for patent-SDG classification using Large Language Models