Minimax rates for cost-sensitive learning on manifolds with approximate nearest neighbours