Exploiting dynamic sparse matrices for performance portable linear algebra operations