Global Convergence of Adjoint-Optimized Neural PDEs