Reduce friction in tabular dataset workflow by eliminating having splits when dataset is loaded

### Feature request

Sorry for cryptic name but I'd like to explain using code itself. When I want to load a specific dataset from a repository (for instance, this: https://huggingface.co/datasets/inria-soda/tabular-benchmark)

```python
from datasets import load_dataset
dataset = load_dataset("inria-soda/tabular-benchmark", data_files=["reg_cat/house_sales.csv"], streaming=True)
print(next(iter(dataset["train"])))
```

`datasets` library is essentially designed for people who'd like to use benchmark datasets on various modalities to fine-tune their models, and these benchmark datasets usually have pre-defined train and test splits. However, for tabular workflows, having train and test splits usually ends up model overfitting to validation split so usually the users would like to do validation techniques like `StratifiedKFoldCrossValidation` or when they tune for hyperparameters they do `GridSearchCrossValidation` so often the behavior is to create their own splits. Even [in this paper](https://hal.archives-ouvertes.fr/hal-03723551) a benchmark is introduced but the split is done by authors.
It's a bit confusing for average tabular user to try and load a dataset and see `"train"` so it would be nice if we would not load dataset into a split called `train `by default.

```diff
from datasets import load_dataset
dataset = load_dataset("inria-soda/tabular-benchmark", data_files=["reg_cat/house_sales.csv"], streaming=True)
-print(next(iter(dataset["train"])))
+print(next(iter(dataset)))
```

### Motivation

I explained it above 😅

### Your contribution

I think this is quite a big change that seems small (e.g. how to determine datasets that will not be load to train split?), it's best if we discuss first!