Transfer updates from BaseDataset and BaseCaption builder to the UC dataset implementations

New parameters for BaseCaptionBuilder:

• n_captions_for_validation: int | str = "all" - how many location caption templates to sample for each validation location (use for loss calculation, note that this is not concept captions). These captions are then embedded and averaged for each location. By default all is very computationally exhaustive (e.g., if you have 1000 templates and 200 locations, that's encoding 200 000 text captions.
• return_aux_ids: bool = False - if caption builder should return column names used in the building of a template. This is important for soft contrastive losses, that treat locations with similar environmental variables as partial positives.

For BaseDataset:

• In _setup() self.df gets subsetted by removing rows for any locations that have missing modality data. Then self.df is converted to records in the BaseDataset.setup()
• New return_name_loc parameter, which helps in debugging, as dataloaders return location ids.
• Cleaner setup for tessera: v1.1 implemented, data is only downloaded if requested, missing tiles can be ignored (removed from self.df)
• Some other minor changes, that depend on when you branched out.

@robknapen @BachirNILU @wkravchyk Please if you can review your BasaDataset subclass implementations.