Cache problem in the `load_dataset` method for local compressed file(s)

## Describe the bug

Cache problem in the `load_dataset` method: when modifying a compressed file in a local folder `load_dataset` doesn't detect the change and load the previous version.

## Steps to reproduce the bug

To test it directly, I have prepared a [Google Colaboratory notebook](https://colab.research.google.com/drive/11Em_Amoc-aPGhSBIkSHU2AvEh24nVayy?usp=sharing) that shows this behavior.

For this example, I have created a toy dataset at: https://huggingface.co/datasets/SaulLu/toy_struc_dataset

This dataset is composed of two versions:
- v1 on commit `a6beb46` which has a single example `{'id': 1, 'value': {'tag': 'a', 'value': 1}}` in file `train.jsonl.gz`
- v2 on commit `e7935f4` (`main` head) which has a single example `{'attr': 1, 'id': 1, 'value': 'a'}` in file `train.jsonl.gz`

With a terminal, we can start to get the v1 version of the dataset
```bash
git lfs install
git clone https://huggingface.co/datasets/SaulLu/toy_struc_dataset
cd toy_struc_dataset
git checkout a6beb46
```
Then we can load it with python and look at the content:
```python
from datasets import load_dataset

path = "/content/toy_struc_dataset"
dataset = load_dataset(path, data_files={"train": "*.jsonl.gz"})
print(dataset["train"][0])
```
Output
```
{'id': 1, 'value': {'tag': 'a', 'value': 1}} # This is the example in v1
```

With a terminal, we can now start to get the v1 version of the dataset
```bash
git checkout main
```
Then we can load it with python and look at the content:
```python
from datasets import load_dataset

path = "/content/toy_struc_dataset"
dataset = load_dataset(path, data_files={"train": "*.jsonl.gz"})
print(dataset["train"][0])
```
Output
```
{'id': 1, 'value': {'tag': 'a', 'value': 1}} # This is the example in v1 (not v2)
```
## Expected results
The last output should have been
```
{"id":1, "value": "a", "attr": 1} # This is the example in v2
```

## Ideas

As discussed offline with Quentin, if the cache hash was ever sensitive to changes in a compressed file we would probably not have the problem anymore.

This situation leads me to suggest 2 other features:
- to also have an `load_from_cache_file` argument in the "load_dataset" method
- to reorganize the cache so that we can delete the caches related to a dataset (cf issue #ToBeFilledSoon)

And thanks again for this great library :hugs:

## Environment info
- `datasets` version: 1.12.1
- Platform: Linux-5.4.104+-x86_64-with-Ubuntu-18.04-bionic
- Python version: 3.7.12
- PyArrow version: 3.0.0