Federated Training of Dual Encoding Models on Small Non-IID Client Datasets