Mitigating Negative Transfer in Multi-Task Learning with Exponential Moving Average Loss Weighting Strategies