Applying map after transform with multiprocessing will cause OOM

### Describe the bug

I have a 30TB dataset. When I perform add_column and cast_column operations on it and then execute a multiprocessing map, it results in an OOM (Out of Memory) error. However, if I skip the add_column and cast_column steps and directly run the map, there is no OOM. After debugging step by step, I found that the OOM is caused at this point, and I suspect it’s because the add_column and cast_column operations are not cached, which causes the entire dataset to be loaded in each subprocess, leading to the OOM. The critical line of code is: https://github.com/huggingface/datasets/blob/e71b0b19d79c7531f9b9bea7c09916b5f6157f42/src/datasets/utils/py_utils.py#L607

Note num_process=1 would not cause OOM. I'm confused.

### Steps to reproduce the bug

For reproduce, you can load dataset and set cache_dir (for caching): amphion/Emilia-Dataset which is a veru large datasets that RAM can not fits.
And apply the map with multiprocessing after a transform operation (e.g. add_column, cast_column).
As long as num_process>1, it must cause OOM.

### Expected behavior

It should not cause OOM.

### Environment info

- `datasets` version: 3.6.0
- Platform: Linux-5.10.134-16.101.al8.x86_64-x86_64-with-glibc2.35
- Python version: 3.10.12
- `huggingface_hub` version: 0.33.1
- PyArrow version: 20.0.0
- Pandas version: 2.3.0
- `fsspec` version: 2024.6.1