Evaluating Performance and Bias of Negative Sampling in Large-Scale Sequential Recommendation Models