Add warning for stop string edge case

[Rocketknight1]

In #31435 @murray-z mentions that the stop string code can throw an error with an unusual stop string: "✿RESULT✿"

The underlying cause is that the ✿ token doesn't exist in the tokenizer vocabulary anywhere. The tokenizer in question has a special-case fallback for unknown characters where it builds them out of other tokens, but the tokens it uses to compose unknown characters have different meanings when considered alone. This means that the stop string code can't figure out how to build the stop string out of the tokenizer vocabulary, because it doesn't have access to the internal special-casing logic.

I can't really see a general solution here without completely rewriting the stop string code to run tokenizer.decode() at each generation step, which would break the ability to compile generation or export it to non-Python environments. Since it's a fairly rare edge case, though, I'm just going to add a proper error message instead!

cc @LysandreJik

[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.

[LysandreJik]

Makes sense, LGTM! Thanks @Rocketknight1