`to_json` in multiprocessing fashion sometimes deadlock

## Describe the bug

`to_json` in multiprocessing fashion sometimes deadlock, instead of raising exceptions. Temporary solution is to see that it deadlocks, and then reduce the number of processes or batch size in order to reduce the memory footprint.

As @lhoestq pointed out, this might be related to https://bugs.python.org/issue22393#msg315684 where `multiprocessing` fails to raise the OOM exception. One suggested alternative is not use `concurrent.futures` instead.

## Steps to reproduce the bug

## Expected results

Script fails when one worker hits OOM, and raise appropriate error.

## Actual results

Deadlock

## Environment info

- `datasets` version: 1.8.1
- Platform: Linux
- Python version: 3.8
- PyArrow version: 6.0.1