AutoTokenizer hash value got change after datasets.map

## Describe the bug
AutoTokenizer hash value got change after datasets.map

## Steps to reproduce the bug
1. trash huggingface datasets cache
2. run the following code:
```python
from transformers import AutoTokenizer, BertTokenizer
from datasets import load_dataset
from datasets.fingerprint import Hasher
tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased')

def tokenize_function(example):
return tokenizer(example["sentence1"], example["sentence2"], truncation=True)

raw_datasets = load_dataset("glue", "mrpc")

print(Hasher.hash(tokenize_function))
print(Hasher.hash(tokenizer))

tokenized_datasets = raw_datasets.map(tokenize_function, batched=True)

print(Hasher.hash(tokenize_function))
print(Hasher.hash(tokenizer))
```
got
```
Reusing dataset glue (/home1/wts/.cache/huggingface/datasets/glue/mrpc/1.0.0/dacbe3125aa31d7f70367a07a8a9e72a5a0bfeb5fc42e75c9db75b96da6053ad)
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 1112.35it/s]
f4976bb4694ebc51
3fca35a1fd4a1251
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 6.96ba/s]
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 15.25ba/s]
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 5.81ba/s]
d32837619b7d7d01
5fd925c82edd62b6
```
3. run raw_datasets.map(tokenize_function, batched=True) again and see some dataset are not using cache.

## Expected results
`AutoTokenizer` work like specific Tokenizer (The hash value don't change after map):
```python
from transformers import AutoTokenizer, BertTokenizer
from datasets import load_dataset
from datasets.fingerprint import Hasher
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')

def tokenize_function(example):
return tokenizer(example["sentence1"], example["sentence2"], truncation=True)

raw_datasets = load_dataset("glue", "mrpc")

print(Hasher.hash(tokenize_function))
print(Hasher.hash(tokenizer))

tokenized_datasets = raw_datasets.map(tokenize_function, batched=True)

print(Hasher.hash(tokenize_function))
print(Hasher.hash(tokenizer))
```

```
Reusing dataset glue (/home1/wts/.cache/huggingface/datasets/glue/mrpc/1.0.0/dacbe3125aa31d7f70367a07a8a9e72a5a0bfeb5fc42e75c9db75b96da6053ad)
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 1091.22it/s]
46d4b31f54153fc7
5b8771afd8d43888
Loading cached processed dataset at /home1/wts/.cache/huggingface/datasets/glue/mrpc/1.0.0/dacbe3125aa31d7f70367a07a8a9e72a5a0bfeb5fc42e75c9db75b96da6053ad/cache-6b07ff82ae9d5c51.arrow
Loading cached processed dataset at /home1/wts/.cache/huggingface/datasets/glue/mrpc/1.0.0/dacbe3125aa31d7f70367a07a8a9e72a5a0bfeb5fc42e75c9db75b96da6053ad/cache-af738a6d84f3864b.arrow
Loading cached processed dataset at /home1/wts/.cache/huggingface/datasets/glue/mrpc/1.0.0/dacbe3125aa31d7f70367a07a8a9e72a5a0bfeb5fc42e75c9db75b96da6053ad/cache-531d2a603ba713c1.arrow
46d4b31f54153fc7
5b8771afd8d43888
```

## Environment info
- `datasets` version: 1.18.0
- Platform: Linux-5.4.0-91-generic-x86_64-with-glibc2.27
- Python version: 3.9.7
- PyArrow version: 6.0.1