[ENH] standardise output format of NBeats and NBeatsKAN estimators for point predictions #1975
Description
Is your feature request related to a problem? Please describe.
As per the findings of #1971, it has been observed that NBeats and NBeatsKAN are non-conformant when it comes to point prediction outputs. Currently, they output a 2d tensor (batch_size, prediction_length), in this case. This needs to be a 3d tensor (batch_size,prediction_length, 1).
Describe the solution you'd like
A simple fix which adds a 3rd dimension to the existing 2D tensor return of the model. If the described fix is not enough, other approaches are welcome. Refer #1966 for some context on the exact output format which is also being tested and currently skipped for NBeats, NBeatsKAN, xLSTMTime and TimeXer (this is being addressed in #1936).
Additional Context
As a simple starting point, here is the code to be referenced for the changes at the link - https://github.com/sktime/pytorch-forecasting/blob/main/pytorch_forecasting/models/nbeats/_nbeats_adapter.py#L69C9-L112C10. Same goes for NBeatsKAN