`_pad_token` attribute?

Thanks for making this open-source! The following function checks for `_pad_token` attribute:
```python
def _tokenize(self, text_sample):
if self.tokenizer._pad_token is None:
# Some tokenizers (e.g. GPT2 tokenizer) have no padding token which causes bugs
raise RuntimeError("If tokenizing on-the-fly, tokenizer must have a pad_token_id")

return self.tokenizer(text_sample["text"], truncation=True, padding="max_length", max_length=self.max_seq_len)
```

But shouldn't it simply check for `pad_token_id`? My tokenizer has `pad_token_id` and `pad_token`, but no `_pad_token`.

Start at the `_tokenize` method shown in the issue and inspect how the tokenizer’s padding attributes are used. Reproduce the case with a tokenizer that has `pad_token_id` and `pad_token` but no `_pad_token`, then check the behavior for GPT2-like tokenizers. Done means valid tokenizers can be tokenized while missing padding support still produces the stated error.

Written by the indexing model from the issue text.