Improve precision of grammar_strict

[shubhamugare]

In some cases, we are underapproximating the next token set, the model may behave weirdly. This is mainly because the grammar-constrained decoding forces the model to generate whitespace " " and the next word separately. The current LLMs are not trained to generate words after whitespace, and the model's quality can degrade.

Hence, in this PR we improve the precision of this approach by using accept sequences of length 3 in certain cases. Mainly, when %ignore tokens such as whitespace are present, this would enable SynCode to have a further lookahead.

Consider a case, input = "I" and the model's next choice is to choose " have" as the next token. SynCode with single lookahead and underapproximating grammar_strict mode, would force the model to generate " ". Typically, during the model's actual training, the model is not trained on inputs ending with whitespace. Thus, in the next step, when input="I ", the model's behavior seems to be poorer.

To fix this, in this PR we allow a longer lookahead accept sequences in some cases.