Scalable Learned Model Soup on a Single GPU: An Efficient Subspace Training Strategy