Allow to give the dataset multiprocessing_context

[ierezell]

Feature request

In Huggingface Trainer, allow to pass the multiprocessing context : https://pytorch.org/docs/stable/data.html#torch.utils.data.DataLoader

Motivation

For a dataset that is loaded on multiple cpu cores, sometimes the fork method creates problems (with polars for example) and the spawn method is more adapted.

Your contribution

I could do a PR. A fix could be to add one more parameter to Trainer and pass it to the Dataloader down the line.

[Rocketknight1]

This looks like a crossover datasets/Trainer issue, so cc @lhoestq @SunMarc @muellerzr