Bayesian Uncertainty for Gradient Aggregation in Multi-Task Learning