Gap-Dependent Bounds for Federated $Q$-learning