Stop Tuning Your Losses: How Uncertainty Can Automatically Balance Multi-Task Learning Models