XLMRobertaTokenizer.detokenize method is not wokring

[onting]

Describe the bug

detokenize method is not available for XLMRobertaTokenizer class

To Reproduce

tokenizer = keras_nlp.models.XLMRobertaTokenizer.from_preset("xlm_roberta_base_multi")
for sentence in tokenizer.detokenize(tokenizer(["한국어 토크나이즈", "This is a bug example"])).numpy():
  print(sentence.decode())

Expected behavior

한국어 토크나이즈
This is a bug example

Additional context

The decode method has been worked with code like below.

def detokenize(self, ids):
        ids = tf.ragged.boolean_mask(ids, tf.not_equal(ids, self.mask_token_id))
        ids = tf.subtract(ids, 1)
        return super().detokenize(ids)

So, it must be related with index mishaps. And it seems like list concatenation mishap like below
https://github.com/keras-team/keras-nlp/blob/2b02bff6d7395898f0ab53476662a972c913f39b/keras_nlp/models/xlm_roberta/xlm_roberta_tokenizer.py#L107-L114
The concatenation increase all indices in previous vocabulary, and it affects on super().detokenize(ids).

[mattdangerw]

Good catch! Valid bug, and indeed it looks like we are just not properly reversing the index remapping that is done after sentepiece tokenization (this is just a weird quirk of xlm roberta).

Are you interested in working on a fix? If not will find an owner soon.

[onting]

@mattdangerw I'm not confident in fixing. Other people should manage this one better.
Thank you.