Adding `sep` to TextConfig

I have a local a `.txt` file that follows the `CONLL2003` format which I need to load using `load_script`. However, by using `sample_by='line'`, one can only split the dataset into lines without splitting each line into columns. Would it be reasonable to add a `sep` argument in combination with `sample_by='paragraph'` to parse a paragraph into an array for each column ? If so, I am happy to contribute!

## Environment
* `python 3.8.10`
* `datasets 2.9.0`

## Snippet of `train.txt`

```txt
Distribution NN O O
and NN O O
dynamics NN O O
of NN O O
electron NN O B-RP
complexes NN O I-RP
in NN O O
cyanobacterial NN O B-R
membranes NN O I-R

The NN O O
occurrence NN O O
of NN O O
prostaglandin NN O B-R
F2α NN O I-R
in NN O O
Pharbitis NN O B-R
seedlings NN O I-R
grown NN O O
under NN O O
short NN O B-P
days NN O I-P
or NN O I-P
days NN O I-P

```

## Current Behaviour
```python
# defining 4 features ['tokens', 'pos_tags', 'chunk_tags', 'ner_tags'] here would fail with `ValueError: Length of names (4) does not match length of arrays (1)`
dataset = datasets.load_dataset(path='text', features=features, data_files={'train': 'train.txt'}, sample_by='line')
dataset['train']['tokens'][0]
>>> 'Distribution\tNN\tO\tO'
```

## Expected Behaviour / Suggestion
```python
# suppose we defined 4 features ['tokens', 'pos_tags', 'chunk_tags', 'ner_tags']
dataset = datasets.load_dataset(path='text', features=features, data_files={'train': 'train.txt'}, sample_by='paragraph', sep='\t')
dataset['train']['tokens'][0]
>>> ['Distribution', 'and', 'dynamics', ... ]

dataset['train']['ner_tags'][0]
>>> ['O', 'O', 'O', ... ]
```