[ENH] Work Items for `ptf-v2`

[phoeenniixx]

We already have a beta version released for ptf-v2, but still a lot of work is to be done.
This issue aims to collect and track the pending work items and improvements that need to be discussed and addressed.

Work Items (not in any specific priority):

• predict function and _pkg wrapper (as mentioned in [API] M layer design #1870 and [STEP] Design Document for Prediction pipeline of ptf-v2 enhancement-proposals#46) implementation
• How to handle categorical variables ?
• Add scalers for preprocessing
• Add more train_test_split strategies
• Add label_encoders to D1 Layer?
• Add support for datasets which cannot fit in memory
• Move the v1 models to v2 (could be of low priority right now)
• Adding Extension Templates for adding new models (and maybe even for datamodules) as mentioned in [ENH] extension templates for neural network classes #1953 (could be of a low priority right now)
• Add MultiLoss support and adapters for nn losses

[phoeenniixx]

FYI @fkiraly, @agobbifbk, @PranavBhatP
Please feel free to add more tasks that I may have missed out

[phoeenniixx]

Also, I was thinking about the datamodules:
Right now, we have 2 - one for ptf native models and one for tslib, so I was thinking, maybe moving forward we should also add a Base Class for these data modules as well?

[fkiraly]

Preliminary thoughts on priorities. I would advise three parallel tracks:

1. v2 feature parity, with priority in decreasing order:

• predict interface
• scalers
• categorical inputs and label encoder

2. adding more models - v1 and/or v2

incl writing an extension template

I think this is important to:

• validate the v2 API as it progresses
• validate the v1 testing framework (and with it v2)
• encourage new contributors

Noting that we will likely have to revisit v2 estimators as we are changing API.

3. attempt full integration with dsipts - using the 2nd party integration pattern as between sktime and prophetverse

[agobbifbk]

My two cents for the numerical scalers and not-in-memory data. We can leverage on partial_fit method of most of the scalers (unfortunately not for categorical variable) without reinventing the wheel
https://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.StandardScaler.html#sklearn.preprocessing.StandardScaler.partial_fit