Concurrency bug when using same cache among several jobs

## Describe the bug
I used to see this bug with an older version of the datasets. It seems to persist.

This is my concrete scenario: I launch several evaluation jobs on a cluster in which I share the file system and I share the cache directory used by huggingface libraries. The evaluation jobs read the same *.csv files. If my jobs get all scheduled pretty much at the same time, there are all kinds of weird concurrency errors. Sometime it crashes silently. This time I got lucky that it crashed with a stack trace that I can share and maybe you get to the bottom of this. If you don't have a similar setup available, it may be hard to reproduce as you really need two jobs accessing the same file at the same time to see this type of bug.

## Steps to reproduce the bug
I'm running a modified version of `run_glue.py` script adapted to my use case. I've seen the same problem when running some glue datasets as well (so it's not specific to loading the datasets from csv files).

## Expected results
No crash, concurrent access to the (intermediate) files just fine.

## Actual results
Crashes due to races/concurrency bugs.

## Environment info

- `datasets` version: 2.3.2
- Platform: Linux-4.18.0-348.23.1.el8_5.x86_64-x86_64-with-glibc2.10
- Python version: 3.8.5
- PyArrow version: 8.0.0
- Pandas version: 1.1.0

Stack trace that I just got with the crash (I've obfuscated some names, it should still be quite informative):

```
Running tokenizer on dataset: 0%| | 0/3 [00:00
main()
File "../../src/models//run_*******.py", line 444, in main
raw_datasets = raw_datasets.map(
File "/*******//envs/tr-crt/lib/python3.8/site-packages/datasets/dataset_dict.py", line 770, in map
{
File "/*******//envs/tr-crt/lib/python3.8/site-packages/datasets/dataset_dict.py", line 771, in
k: dataset.map(
File "/*******//envs/tr-crt/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 2376, in map
return self._map_single(
File "/*******/envs/tr-crt/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 551, in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
File "/*******//envs/tr-crt/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 518, in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
File "/*******/envs/tr-crt/lib/python3.8/site-packages/datasets/fingerprint.py", line 458, in wrapper
out = func(self, *args, **kwargs)
File "/*******//envs/tr-crt/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 2776, in _map_single
buf_writer, writer, tmp_file = init_buffer_and_writer()
File "/*******//envs/tr-crt/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 2696, in init_buffer_and_writer
tmp_file = tempfile.NamedTemporaryFile("wb", dir=os.path.dirname(cache_file_name), delete=False)
File "/*******//envs/tr-crt/lib/python3.8/tempfile.py", line 541, in NamedTemporaryFile
(fd, name) = _mkstemp_inner(dir, prefix, suffix, flags, output_type)
File "/*******//envs/tr-crt/lib/python3.8/tempfile.py", line 250, in _mkstemp_inner
fd = _os.open(file, flags, 0o600)
FileNotFoundError: [Errno 2] No such file or directory: '/*******/cache-transformers//transformers/csv/default-ef9cd184210742a7/0.0.0/51cce309a08df9c4d82ffd9363bbe090bf173197fc01a71b034e8594995a1a58/tmps8l6j5yc'
```

As I ran 100s of experiments last year for an empirical paper, I ran into this type of bugs several times. I found several bandaid/work-arounds, e.g., run one job first that caches the dataset => eliminate concurrency; OR use unique caches => eliminate concurrency (but increase storage space), etc. and it all works fine.

I'd like to help you fixing this bug as it's really annoying to always apply the work arounds. Let me know what other info from my side could help you figure out the issue.

Thanks for your help!