Learning from Similar Linear Representations: Adaptivity, Minimaxity, and Robustness