Token Type IDs returned from the tokenizer for T5 don't work with special tokens

With `transformers-3.3.1`:

```
import transformers
t = transformers.AutoTokenizer.from_pretrained('t5-small')
t.encode_plus(["a"], ["b"], add_special_tokens=True, return_token_type_ids=True)
```

This results in
```
{'input_ids': [9, 1, 115, 1], 'token_type_ids': [0, 1], 'attention_mask': [1, 1, 1, 1]}
```

As you can see, the token type IDs don't align with the other outputs.