Extend GELOS DataSet and Modules for Fire

[Ruphai]

Added Extension to load the wildfire datasets in the GELOSDataset and DataModule.

[Ruphai]

@DWGodwin I added additional changes to be able to load the fire mask alongside the other chips. To do this, I added a boolean to check if it is fire or not (something like a yearly_time_series or change_detection might be more descriptive for adding new datasets like it.

[Ruphai]

An adjacent issue, I noticed that the dataset module loads the images following the order ofT, H, W, C to be able to pad the inputs to 98X98 using Albumentations for the Prithvi 600M Embedding generation pipeline. But for the embedding generation, Terratorch expects the input image to follow the order of C, T, H, W.
Could you clarify where in the pipeline the switch between T, H, W, C and C, T, H, W needs to be made for the embedding generation run as expected?

[DWGodwin]

Apologies for not seeing this sooner. This is handled by the custom terratorch transform UnflattenTemporalFromChannels, which takes flattened tensors and returns them in format C, T, H, W. The function combines the separate concerns of unflattening and transposition.