Make llama_samplers.cpp throw an error/exception on lack of choices during sampling

[SteelPh0enix]

I've recently noticed, while testing samplers options in llama.cpp, that current implementation aborts and crashes when sampling code (specifically, llama_sampler_softmax_impl in my case) encounters a situation where the length of tokens array is 0.

IMO this should not be a critical error of this magnitude, and instead it should exit graciously, with either an exception or some other kind of error-propagation mechanism. However, i'm currently unfamiliar with llama.cpp codebase so i don't know what kind of approach would be preferred, hence i started this discussion.

[ggerganov]

Open an issue with repro steps and this will get fixed.