Update pytorch lightning import to work with Ray and other libraries.

[ru-ko]

Is your feature request related to a current problem? Please describe.
Importing and using lightning.pytorch and pytorch_lightning in the same program causes conflicts. Ray has migrated their tuning to ray.tune.Tuner and is deprecating ray.tune.run. Ray uses a conditional import, importing pytorch_lightning only if the other is not available. However, it will import lightning.pytorch first, where the darts pl_forecasting_module imports pytorch_lightning. This may also apply to other places.

Describe proposed solution
Darts should incorporate a conditional import that matches this priority in pl_forecasting_module, lr_finder, etc as follows:

#import pytorch_lightning as pl
def import_lightning():  # noqa: F402
    try:
        import lightning.pytorch as pl
    except ModuleNotFoundError:
        import pytorch_lightning as pl
    return pl

pl = import_lightning()

Note that files may import more than just pl - those other imports could also be wrapped under the conditional.

Describe potential alternatives
Change imports of pytorch_lightning to lightning.pytorch

Additional context
I noticed this issue when attempting to use ray.tune.Tuner and ray.train.torch.TorchTrainer in conjunction with RNNModel. I solved it by making the change in the proposed solution.

[madtoinou]

Hi @ru-ko,

We are aware of this problem, it's already tracked by #2198.

I like the approach you're proposing, would you be willing to open a PR to implement it and see if it solves the problem?

[ru-ko]

Ope, didn't realize it was tracked by that. I'll start a branch for it and make changes when I can, although it'll be slow going.