Clarify byte-level pretokenizer behavior

[michaelalbada]

Clarifies the ByteLevel pre-tokenizer docs around raw byte input.

ByteLevel operates on Unicode strings and maps their UTF-8 bytes to visible byte-level characters. Passing arbitrary bytes decoded as Latin-1 can therefore re-encode bytes above 127 as UTF-8 and produce unexpected tokens.

This PR:

• documents that limitation in the ByteLevel docs
• adds a short raw-byte encoding example using the GPT-2 byte-to-Unicode mapping
• explains that already-mapped byte-level text should skip normalization and pre-tokenization

Designed to address [[https://github.com//issues/1877]]

[michaelalbada]

Hi @Pierrci, would you mind taking a look at this PR?

[ArthurZucker]

ty for the writeup! we are kinda gonna remove it soon, as performance wise this really dose not make a lot of sense to have!