Impossible to save a mapped dataset to disk

### Describe the bug

I want to play around with different hyperparameters when training but don't want to re-map my dataset with 3 million samples each time for tens of hours when I [fully fine-tune SDXL](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image_sdxl.py).

After I do the mapping like this:
```
train_dataset = train_dataset.map(compute_embeddings_fn, batched=True)
train_dataset = train_dataset.map(
compute_vae_encodings_fn,
batched=True,
batch_size=16,
)
```
and try to save it like this:
`train_dataset.save_to_disk("test")`
i get this error ([full traceback](https://pastebin.com/kq3vt739)):
```
TypeError: Object of type function is not JSON serializable
The format kwargs must be JSON serializable, but key 'transform' isn't.
```

But what is interesting is that pushing to hub works like that:
`train_dataset.push_to_hub("kopyl/mapped-833-icons-sdxl-1024-dataset", token=True)`
Here is the link of the pushed dataset: https://huggingface.co/datasets/kopyl/mapped-833-icons-sdxl-1024-dataset

### Steps to reproduce the bug

Here is the self-contained notebook:

https://colab.research.google.com/drive/1RtCsEMVcwWcMwlWURk_cj_9xUBHz065M?usp=sharing

### Expected behavior

It should be easily saved to disk

### Environment info

NVIDIA A100, Linux (NC24ads A100 v4 from Azure), CUDA 12.2.

[pip freeze](https://pastebin.com/QTNb6iru)