Fix XLM-RoBERTa detokenize()

[abheesht17]

Resolves #1282

Tested manually

import io
import os

import sentencepiece
import tensorflow as tf
from keras_nlp.backend import keras

from keras_nlp.models.xlm_roberta.xlm_roberta_tokenizer import (
    XLMRobertaTokenizer,
)

bytes_io = io.BytesIO()
vocab_data = tf.data.Dataset.from_tensor_slices(
    ["the quick brown fox", "the earth is round"]
)
sentencepiece.SentencePieceTrainer.train(
    sentence_iterator=vocab_data.as_numpy_iterator(),
    model_writer=bytes_io,
    vocab_size=10,
    model_type="WORD",
    unk_id=0,
    bos_id=1,
    eos_id=2,
)
proto = bytes_io.getvalue()

tokenizer = XLMRobertaTokenizer(proto=proto)

input_data = tf.constant([[4, 9, 5, 7]])
output = tokenizer.detokenize(input_data)
output

Output:

<tf.Tensor: shape=(1,), dtype=string, numpy=array([b'the quick brown fox'], dtype=object)>

[abheesht17]

/gcbrun

[mattdangerw]

Looks good! I will add bad the detokenize tests while we are at it.