Recon: Reducing Conflicting Gradients from the Root for Multi-Task Learning