Distribution-Dependent Rates for Multi-Distribution Learning