Rethink the "splits / mode" behaviour

Ideas:

- have a train and an eval mode, like now
- Ability to specify multiple splits for each mode. Run the splits sequentially (makes sense for eval, maybe less for train. Option to merge them in common batches, or run batches in alternance ?)
- Metrics should use the {split} and not the {mode} to save their data ?

Use cases can be train/val/test, or multi source datasets (example for data augmentation with another dataset in train).
Or having multiple eval dataset, with different metrics

No files, tests, or entry points are named. First locate the current train/eval split handling and metric data storage, then resolve whether multiple splits run sequentially, merge into batches, or alternate before defining implementation and tests for the agreed behavior.

Written by the indexing model from the issue text.