Revisiting differentially private linear regression: optimal and adaptive prediction & estimation in unbounded domain