Probabilistic High Memory Usage and Freeze on Python 3.10

### Describe the bug

A probabilistic issue encountered when processing datasets containing PIL.Image columns using the huggingface/datasets library on Python 3.10. The process occasionally experiences a sudden and significant memory spike, reaching 100% utilization, leading to a complete freeze. During this freeze, the process becomes unresponsive, cannot be forcefully terminated, and does not throw any exceptions.

I have attempted to mitigate this issue by setting `datasets.config.IN_MEMORY_MAX_SIZE`, but it had no effect. In fact, based on the document of `load_dataset`, I suspect that setting `IN_MEMORY_MAX_SIZE` might even have a counterproductive effect.

This bug is not consistently reproducible, but its occurrence rate significantly decreases or disappears entirely when upgrading Python to version 3.11 or higher. Therefore, this issue also serves to share a potential solution for others who might encounter similar problems.

### Steps to reproduce the bug

Due to the probabilistic nature of this bug, consistent reproduction cannot be guaranteed for every run. However, in my environment, processing large datasets like timm/imagenet-1k-wds(whether reading, casting, or mapping operations) almost certainly triggers the issue at some point.
The probability of the issue occurring drastically increases when num_proc is set to a value greater than 1 during operations.

When the issue occurs, my system logs repeatedly show the following warnings:
```
WARN: very high memory utilization: 57.74GiB / 57.74GiB (100 %)
WARN: container is unhealthy: triggered memory limits (OOM)
WARN: container is unhealthy: triggered memory limits (OOM)
WARN: container is unhealthy: triggered memory limits (OOM)
```

### Expected behavior

The dataset should be read and processed normally without memory exhaustion or freezing. If an unrecoverable error occurs, an appropriate exception should be raised.

I have found that upgrading Python to version 3.11 or above completely resolves this issue. On Python 3.11, when memory usage approaches 100%, it suddenly drops before slowly increasing again. I suspect this behavior is due to an expected memory management action, possibly involving writing to disk cache, which prevents the complete freeze observed in Python 3.10.

### Environment info

- `datasets` version: 4.0.0
- Platform: Linux-5.15.0-71-generic-x86_64-with-glibc2.35
- Python version: 3.10.12
- `huggingface_hub` version: 0.33.4
- PyArrow version: 20.0.0
- Pandas version: 2.3.1
- `fsspec` version: 2025.3.0