Support end/beginning of sequence other than BOS and EOS

[rlouf]

This issue is very general; I do not have a MRE yet.

Large Language Models use the EOS token to signify that the generation is over. Instruct models are different: they have other special tokens to mark the beginning and end of an instruction, as well as the beginning and end of the model's response, see for instance OpenAI's. My impression with the current setup is that generation does not stop when these special tokens are generated, and that generating a turn start token is not disallowed, unlike with the BOS and EOS token.

I noticed this when I went over the SCP examples. There we use JSON-based structured generation to generate the different entries. It turns out that some of the JSON fields contained these special tokens, and then more text, as you can see in this commit where I made the correction by hand.