Fix bytelevel decode of added tokens + 27x faster deserialization

[ArthurZucker]

1. When adding the token we should normalize it if normalize
2. Change Ahocorasic for daahorse
3. Change the algo to simplify.
4. Add python test for ByteLevel decode with normalizer

Todo: check if normalizer ByteLevel vs PreTokenizer which is better?
mains vs this branch

non special deserialize_added_vocab_100000_norm_none
                        time:   [7.5797 s 7.6353 s 7.6861 s]
                        change: [+2386.1% +2424.7% +2460.7%] (p = 0.00 < 0.10)
                        Performance has regressed.

daahorse just insane here.

gains mostly on non-special (probably because default is non special == normalize)

Also previously every add_tokens call renormalized the entire vocab to build the regex, now it only normalizes what's new. This can be wrong if normalizer changes tho. We could "guard" against this by updating tokenizer.normalizer setter to refresh added tokens.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[ArthurZucker]

the 1st has normalized false -> broken but the next 2 are properly re-constructed