On the Linear Speedup of Personalized Federated Reinforcement Learning with Shared Representations