On the linearity of large non-linear models: when and why the tangent kernel is constant