Federated Reinforcement Learning: Linear Speedup Under Markovian Sampling