Currently, forward of tslib models returns 4D tensors in case of QuantileLoss and 2D(for single-target, point loss in DLinear) and 3D(for multi-target, point loss in DLinear and for point loss in TimeXer).

We should standardize the output return of all models to 3D tensors (for single-target) and list of 3D tensors (for multi-target), like other models in pytorch-forecasting