Training Deep Learning Models with Norm-Constrained LMOs