Batch inference for HF models called with an array of prompts

[rlouf]

The generation method in hf_transformers.py is decorated with outlines.vectorize. Thus when we pass, say a 2-dimensional array, we run the model twice when it would be more efficient to flatten the input array and perform batch inference. We thus need to implement a mechanism that allows to flatten the array before passing it to the decorated function.

To do this, we have a choice between implementing a new decorator (preferred) or change the behavior of outlines.vectorize.

[rlouf]

This is being fixed in #139