[ENH] Add device parameter to classes.

[fnhirwa]

• PyTorch-Forecasting version: 1.0.0
• PyTorch version: 2.4.0
• Python version: 3.9.19
• Operating System: MacOs(Darwin)

Expected behavior

The current implementation of different classes seems to be using devices in a fashion that I can say is complex and users have no control exactly on how they can easily switch between devices.

https://github.com/jdb78/pytorch-forecasting/blob/81aee6650ed3de0c3071c9ce1fce19eec7fc24a7/pytorch_forecasting/metrics/distributions.py#L470

Like here the device being used is the one of the input which is somehow complex for the interface being used.

I suggest the Introduction of the device parameter to the classes. This would give users control over the device control and switching, rather than relying entirely on input data devices.

[XinyuWuu]

The distribution is implemented by lightning. I am not sure if we can set it manually.

import lightning.pytorch as pl
trainer = pl.Trainer(
    accelerator="cpu",
)

[fnhirwa]

I think we should find a better way to handle the accelerator case as I don't think it would be sensible to set these for users: https://github.com/jdb78/pytorch-forecasting/blob/bb6c8a2243c35ca35c2c0e14093d352430fee6d0/pytorch_forecasting/models/temporal_fusion_transformer/tuning.py#L177

We can pass it as some optional parameter.

[fkiraly]

This is the tuning code, I think it's extraneous to the models - as long as the models themselves have it not hard-coded, I consider this less of a problem. I feel this one tuning routine does a lot of hard-coding and things extraneous to an otherwise very consistent architectural design.