No-Regret Learning with Unbounded Losses: The Case of Logarithmic Pooling