Stopping Rules for Stochastic Gradient Descent via Anytime-Valid Confidence Sequences