Allow passing a multiprocessing context to functions that support `num_proc`

### Feature request

Allow specifying [a multiprocessing context](https://docs.python.org/3/library/multiprocessing.html#contexts-and-start-methods) to functions that support `num_proc` or use multiprocessing pools. For example, the following could be done:

```python
dataset = dataset.map(_func, num_proc=2, mp_context=multiprocess.get_context("spawn"))
```

Or at least the multiprocessing start method ("fork", "spawn", "fork_server" or `None`):

```python
dataset = dataset.map(_func, num_proc=2, mp_start_method="spawn")
```

Another option could be passing the `pool` as an argument.

### Motivation

By default, `multiprocess` (the `multiprocessing`-fork library that this repo uses) uses the "fork" start method for multiprocessing pools (for the default context). It could be changed by using `set_start_method`. However, this conditions the multiprocessing start method from all processing in a Python program that uses the default context, because [you can't call that function more than once](https://docs.python.org/3/library/multiprocessing.html#contexts-and-start-methods:~:text=set_start_method()%20should%20not%20be%20used%20more%20than%20once%20in%20the%20program.). My proposal is to allow using a different multiprocessing context, not to condition the whole Python program.

One reason to change the start method is that "fork" (the default) makes child processes likely deadlock if thread pools were created before (and also this is not supported by POSIX). For example, this happens when using PyTorch because OpenMP threads are used for CPU intra-op parallelism, which is enabled by default (e.g., for context see [`torch.set_num_threads`](https://pytorch.org/docs/stable/generated/torch.set_num_threads.html)). This can also be fixed by setting `torch.set_num_threads(1)` (or similarly by other methods) but this conditions the whole Python program as it can only be set once to guarantee its behavior (similarly to). There are noticeable performance differences when setting this number to 1 even when using GPU(s). Using, e.g., a "spawn" start method would solve this issue.

For more context, see:

* https://discuss.huggingface.co/t/dataset-map-stuck-with-torch-set-num-threads-set-to-2-or-larger/37984
* https://discuss.huggingface.co/t/using-num-proc-1-in-dataset-map-hangs/44310

### Your contribution

I'd be happy to review a PR that makes such a change. And if you really don't have the bandwidth for it, I'd consider creating one.