Add support for dataloader samplers

[nkaenzig]

Closes #712

• Adds a samplers argument to eva.DataModule, so we can enable custom samplers in the dataloaders.
  • This enables for instance label efficiency experiments, as we can reduce the number of training samples for the existing downstream tasks.
• Adds a BalancedSampler which supports balanced class sample data loading for classification tasks.

How to use

Just add this to the init_args of the DataModule in your yaml config:

    samplers:
      train:
        class_path: eva.core.data.samplers.classification.BalancedSampler
        init_args:
          num_samples: 10

(For online mode, specify the sampler in samplers.train:, while for offline mode in samplers.predict:).

Make sure that shuffle: false for the dataloader config of the corresponding split.