Global Convergence of MAML and Theory-Inspired Neural Architecture Search for Few-Shot Learning