Revealing the influence of participant failures on model quality in cross-silo Federated Learning