Adding new splits to a dataset script with existing old splits info in metadata's `dataset_info` fails

### Describe the bug

If you first create a custom dataset with a specific set of splits, generate metadata with `datasets-cli test ... --save_info`, then change your script to include more splits, it fails.

That's what happened in https://huggingface.co/datasets/mrdbourke/food_vision_199_classes/discussions/2#6385fd1269634850f8ddff48.

### Steps to reproduce the bug

1. create a dataset with a custom split that returns, for example, only `"train"` split in `_splits_generators'`. specifically, if really want to reproduce, copy `https://huggingface.co/datasets/mrdbourke/food_vision_199_classes/blob/main/food_vision_199_classes.py
2. run `datasets-cli test dataset_script.py --save_info --all_configs` - this would generate metadata yaml in `README.md` that would contain info about splits, for example, like this:
```
splits:
- name: train
num_bytes: 2973286
num_examples: 19747
```
3. make changes to your script so that it returns another set of splits, for example, `"train"` and `"test"` (uncomment [these lines](https://huggingface.co/datasets/mrdbourke/food_vision_199_classes/blob/main/food_vision_199_classes.py#L271))
4. run `load_dataset` and get the following error:
```python
Traceback (most recent call last):
File "/home/daniel/code/pytorch/env/bin/datasets-cli", line 8, in
sys.exit(main())
File "/home/daniel/code/pytorch/env/lib/python3.8/site-packages/datasets/commands/datasets_cli.py", line 39, in main
service.run()
File "/home/daniel/code/pytorch/env/lib/python3.8/site-packages/datasets/commands/test.py", line 141, in run
builder.download_and_prepare(
File "/home/daniel/code/pytorch/env/lib/python3.8/site-packages/datasets/builder.py", line 822, in download_and_prepare
self._download_and_prepare(
File "/home/daniel/code/pytorch/env/lib/python3.8/site-packages/datasets/builder.py", line 1555, in _download_and_prepare
super()._download_and_prepare(
File "/home/daniel/code/pytorch/env/lib/python3.8/site-packages/datasets/builder.py", line 913, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "/home/daniel/code/pytorch/env/lib/python3.8/site-packages/datasets/builder.py", line 1356, in _prepare_split
split_info = self.info.splits[split_generator.name]
File "/home/daniel/code/pytorch/env/lib/python3.8/site-packages/datasets/splits.py", line 525, in __getitem__
instructions = make_file_instructions(
File "/home/daniel/code/pytorch/env/lib/python3.8/site-packages/datasets/arrow_reader.py", line 111, in make_file_instructions
name2filenames = {
File "/home/daniel/code/pytorch/env/lib/python3.8/site-packages/datasets/arrow_reader.py", line 112, in
info.name: filenames_for_dataset_split(
File "/home/daniel/code/pytorch/env/lib/python3.8/site-packages/datasets/naming.py", line 78, in filenames_for_dataset_split
prefix = filename_prefix_for_split(dataset_name, split)
File "/home/daniel/code/pytorch/env/lib/python3.8/site-packages/datasets/naming.py", line 57, in filename_prefix_for_split
if os.path.basename(name) != name:
File "/home/daniel/code/pytorch/env/lib/python3.8/posixpath.py", line 143, in basename
p = os.fspath(p)
TypeError: expected str, bytes or os.PathLike object, not NoneType
```
5. bonus: try to regenerate metadata in `README.md` with `datasets-cli` as in step 2 and get the same error.

This is because `dataset.info.splits` contains only `"train"` split so when we are doing `self.info.splits[split_generator.name]` it tries to infer smth like `info.splits['train[50%]']` and that's not the case and it fails.

### Expected behavior

to be discussed?

This can be solved by removing splits information from metadata file first. But I wonder if there is a better way.

### Environment info

- Datasets version: 2.7.1
- Python version: 3.8.13