Interpretable Anomaly Detection in the LHC Main Dipole Circuits With Nonnegative Matrix Factorization