Geometric Insights into the convergence of Non-linear TD Learning