Optimal Semi-supervised Estimation and Inference for High-dimensional Linear Regression