Towards Explainable Adjudicative Variance: Quantifying Judicial Discretion via Gated Multi-Task Learning