De-biasing the Lasso: Optimal Sample Size for Gaussian Designs