Add record_datasets_as_runs setting to trainer

[nkaenzig]

Summary

• Adds a record_datasets_as_runs parameter to Trainer for evaluation sessions with multiple val and/or test datasets.
• When enabled, each dataset output is recorded as a separate run in the session summary instead of being kept as separate datasets within a single run.
• Passes the setting through the Interface CLI entry points and run_evaluation_session.

How to use

Set record_datasets_as_runs: true in the trainer config, and provide a list of datasets in the val or test section of the data config. Each dataset entry can point to a different data path, for example predictions from different offline runs:

trainer:
  class_path: eva.Trainer
  init_args:
    record_datasets_as_runs: true
    # ... other trainer args

data:
  class_path: eva.DataModule
  init_args:
    datasets:
      val:
        - class_path: eva.language.datasets.TextPredictionDataset
          init_args:
            path: ./predictions/run_1/manifest.jsonl
            split: val
        - class_path: eva.language.datasets.TextPredictionDataset
          init_args:
            path: ./predictions/run_2/manifest.jsonl
            split: val
        - class_path: eva.language.datasets.TextPredictionDataset
          init_args:
            path: ./predictions/run_3/manifest.jsonl
            split: val

In the generated results.json, metrics are reported as one dataset entry whose values list contains the per-dataset run values together with the aggregated mean and stdev:

{
  "metrics": {
    "val": [
      {
        "val/AverageLoss": {
          "mean": 1.58,
          "stdev": 0.0,
          "values": [1.60, 1.55, 1.59]
        },
        "val/BinaryBalancedAccuracy": {
          "mean": 0.50,
          "stdev": 0.01,
          "values": [0.49, 0.51, 0.50]
        }
      }
    ]
  }
}

[DriessenA]

@nkaenzig Any chance we can merge this PR? I'd be happy to help if needed :)

[MaxFeucht]

LGTM