Add support for reasoning models

[RobinPicard]

Expose 2 new keywords for generation:

• end_thinking_tag: a string indicating the tag used by the reasoning model to indicate that thinking is finished (and so that we should start constraining the generation)
• thinking_max_tokens: an int giving the maximum number of tokens during which the model can think, after that number is reached, we force the generation of the end of thinking token

Not supported:

• Models for which the end of the thinking does not correspond to a single token

If we want to capture the content of the thinking in the future when we will return an object with various attributes instead of just the text output, we could add an argument start_thinking_tag for the models that use one.

[rlouf]

Suggested change

	The maximum number of tokens the model can think about. Only used when
	The maximum number of tokens the model can think for. Only used when

[rlouf]

Isn't it possible to only build a specialized logits processor that the backends are unaware of? You should be able to not call the logits biasing function as long as </think> has not been generated, and limit the number of tokens from the logits processor.

[RobinPicard]

Initially I wanted to wrap the logits processor into another one that would just not bias anything until we encounter the token and then it calls the other tokenizer it wraps, the problem is that it does not work for batching as the different sequences may not all stop thinking at the same time.