FLuID: Mitigating Stragglers in Federated Learning using Invariant Dropout