Learning Latent Architectural Distribution in Differentiable Neural Architecture Search via Variational Information Maximization