This PR simplifies the dataloader and input-data creation.
New features

• create_zarr_archive.py now generates one huge zarr-archive directly from the COSMO-2 GRIB-files. @cosunae this might be of interest for offline data preparation.
• weather_dataset.py now loads one zarr archive and __get_item__ now also returns the datetimes of the current batch
• In ar_model.py the datetimes of the batch are compared to constants.EVAL_DATETIMES. Massively simplifying model testing and prediction @clechartre this is certainly relevant for prediction/verification.

Reasoning

• After discussions within ECMWF/Neural-LAM it became pretty clear that one single zarr is the way to go. Lazy loading and parallelization allow for huge archives. The reason why it didn't work for me last year was most likely related to the issues we had with /scratch being extremly slow and unresponsive. It's also much easier to share dataloaders and datasets this way.
• Verification of the model is much simplified keeping track of the datetime of each batch. Model predictions can still be easily exported using the constants.STORE_EXAMPLE_DATA.

Notes

• Zarr creation is ongoing, it roughly takes 3-4 days without Balfrin crashing.
• For now I have created a small cosmo_single dataset that can be used for training and evaluation.
• A new dummy example.ckpt was uploaded for evaluation/testing (trained on very small dataset)
• Maybe it makes sense to wait for the full zarr creation and retraining of the model before this PR is merged. I really wanted to share the code already to prevent duplicated work though.