Support for negative n_best when sampling

[LaurentMazare]

Thanks for these bindings, we've been using them a lot to train our models and have been very happy with them. One thing though is that negative values for n_best when using sampling doesn't seem to be supported, e.g. here the parameter is a size_t that got cast to an int. However on the C++ side it seems that negative values can be used to sample across all possibilities, see this documentation. Is there already a way to do this on the rust side? If not do you think it would be possible to allow negative values to match the C++ api? Happy to make a PR if that helps.

[danieldk]

Sorry, I completely missed this issue. Adding support sounds great! It seems that what negative value is used doesn't matter:

https://github.com/google/sentencepiece/blob/273449044caa593c2fd7eb7550cb3ab2cff93f1a/src/sentencepiece_processor.cc#L689

So maybe it's best to introduce an enum for this argument?