Federated Stochastic Approximation under Markov Noise and Heterogeneity: Applications in Reinforcement Learning