Get nowcasting_dataset running on OCF's on-premises server

[JackKelly]

The source Zarr data is stored locally.

To start with, I'll copy and modify scripts/prepare_ml_training_data.py to create prepare_ml_training_data_on_premises.py, and then, depending on where we land with that, I'll maybe try to merge the changes into a single prepare_ml_training_data.py which works both locally and in the cloud :)