Fixing camembert tokenization

[thomwolf]

The original fairseq implmentation of Camembert has a bunch of duplicate tokens in the dictionary, in particular there are two <unk> tokens but only the index of the first <unk> should be used:

import torch
camembert = torch.hub.load('pytorch/fairseq', 'camembert.v0')
list(camembert.task.source_dictionary[i] for i in range(10))
>>> ['<s>', '<pad>', '</s>', '<unk>', '<unk>', '<s>', '</s>', ',', '▁de', '.']

This PR updates Camembert tokenizer to fix this behavior and as a consequence fixes #2019 and #2020

[thomwolf]

Merging now to fix the xlnet test issue on master at the same time.

[julien-c]

Also cc'ing @louismartin on this.

[louismartin]

Thanks for fixing that.
This comes from a problem in fairseq where special tokens are added twice when using SentencePiece.
Cross-referencing the fairseq issue: facebookresearch/fairseq#1309