Convergence of Gradient Descent for General Neural Network Architectures Beyond the NTK Regime