[Tokenizer Serialization] Fix the broken serialisation

[ArthurZucker]

What does this PR do?

Should fix some serialization issues, mostly save_pretrained with all the init kwargs, and from_pretrained with dicts.
fixes #26732

With main:

>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("EleutherAI/Llemma_7b", use_fast=False)
File ~/Work/transformers/src/transformers/tokenization_utils_base.py:2253, in PreTrainedTokenizerBase._from_pretrained(cls, resolved_vocab_files, pretrained_model_name_or_path, init_configuration, token, cache_dir, local_files_only, _commit_hash, _is_local, *init_inputs, **kwargs)
   2251     if added_tokens_map != {} and init_kwargs[key] is not None:
   2252         if key != "additional_special_tokens":
-> 2253             init_kwargs[key] = added_tokens_map.get(init_kwargs[key], init_kwargs[key])
   2255 init_kwargs["added_tokens_decoder"] = added_tokens_decoder
   2256 # convert {'__type': 'AddedToken', 'content': '<ent>', 'lstrip': False, 'normalized': True, ...} to AddedTokens

TypeError: unhashable type: 'dict'

This is because the tokenizer had special tokens saved as dicts, and the call to convert_added_tokens. is made after this.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[LysandreJik]

Thank you @ArthurZucker

[ArthurZucker]

Pegasus is the only slow failure I witnessed so checking this now before merging!

[ArthurZucker]

Ok, the issue is that when we force the added tokens encoder in the slow tokenizer, the fast of course can't do this. So the eos token gets replaced at index 0 in slow but not in fast.
Will update to force the default vocab to the default tokens.