A Matrix Factorization Approach for Learning Semidefinite-Representable Regularizers