Improve the algorithm for single-token generation

[Francesco215]

As of now, the code of model.generate() doesn't cache the value of each node of the model, therefore it has to re-calcutate everyting when generating each new token