Map cache with tokenizer

Similar issue to https://github.com/huggingface/datasets/issues/5985, but across different sessions rather than two calls in the same session.

Unlike that issue, explicitly calling tokenizer(my_args) before the map() doesn't help, because the tokenizer was created with a different hash to begin with...

setup
```
from transformers import AutoTokenizer
AutoTokenizer.from_pretrained('bert-base-uncased').save_pretrained("tok")
```

this prints different value each time
```
from transformers import AutoTokenizer
from datasets.utils.py_utils import dumps # Huggingface datasets
print(hash(dumps(AutoTokenizer.from_pretrained("tok"))))
```