The Power and Limitation of Pretraining-Finetuning for Linear Regression under Covariate Shift