The Sample-Communication Complexity Trade-off in Federated Q-Learning