Feature/add special tokens

[travismorton]

This allows a pretrained tokenizer to be loaded with additional special tokens. It enables the creation of a Whisper tokenizer which is based on the GPT2 tokenizer with new special tokens. Usage looks like:

{:ok, tokenizer} = Tokenizers.Tokenizer.from_pretrained("gpt2", additional_special_tokens: ["<|special|>"])
# or directly from file
{:ok, tokenizer} = Tokenizers.Tokenizer.from_file("tokenizer.json", additional_special_tokens: ["<|special|>"])

Full Whisper tokenizer example here

I'm not sure what the preferred way of passing opts through to the tokenizer is, interested in feedback there. I left it open to have more options in the future for possibly adding (non-special) tokens and other values that could modify the initial state.

[josevalim]

💚 💙 💜 💛 ❤️