Replicable Learning of Large-Margin Halfspaces