[WIP] Generic multimodal dataset and MultiMAE

[blumenstiel]

This pr includes:

• A generic multimodal dataset that can process multiple image and non-image modalities (GenericMultiModalDataModule)
• MultiMAE backbone (multimae_base and multimae_large, no pre-trained models)
• The generic dataset comes versions for segmentation, pixel-wise regression, and scalar tasks
• The data config is based on the generic datasets. But instead of passing one train_data_root, you can now pass a directory with the modalities as keys.

    train_data_root:
      S2L2A: data/sen1floods11/data/data/flood_events/HandLabeled/S2L2AHand
      S1: data/sen1floods11/data/data/flood_events/HandLabeled/S1Hand
      LULC: data/sen1floods11/data/data/flood_events/HandLabeled/LULCHand

• Additional parameters are:
  • modalities (list of all modality names)
  • image_modalities (list of image modalities, optional, otherwise considers all modalities as images)
  • rgb_modality, task (e.g. segmentation, pixel-wise regression, scalar (similar to classification etc.))
  • shared_transforms (apply the same transformation to all images, default True)
  • chunk_data (if the data already includes a sample dimension, default False)
  • sample_num_modalities (int, sample a subset of the modalities per batch, default None)
  • sample_replace (sample modalities with replacement, default False)
  • concat_bands (concatenate all images modalities along the band dim which enables running single-modal models ,default False)

Example configs: multimae_sen1floods11.yaml multimodal_prithvi_sen1floods11.yaml

Open Todos:

• Implement allow_missing_modalities on a data module level. Idea: Add nan data for missing modalities.
• Update docs
• Check plotting code for scalar tasks
• Add pre-trained RGB MultiMAE model and weight loading
• Add tests