Nonlinear SVD with Asymmetric Kernels: feature learning and asymmetric Nystr\"om method