Implement a new Decoder to support Byte->Char hack spm conversion.

[Narsil]

Has to go after: #872

And implement this algorithm:

https://github.com/huggingface/tokenizers/pull/896/files#diff-0dce407dfb4cba577e8da3aecb16d5e52d0e35faa1df70d1845c69790ca1651cR81

Basically, when decoding a token that "looks" like a byte <0x16> then, treat it as a byte.
This trick is used in SPM too and makes it possible to reduce the vocabulary size by not including every unicode codepoint in the tokenizer, without actually reworking everything to iterate on pure bytes.

[Narsil]

Closing this, as the encoding part is not being done, and ByteLevel already implements it.