Easier perplexity computation

# 🚀 Feature request

The docs provide a method to evaluate perplexity for a GPT-2 model, one example at a time (https://huggingface.co/transformers/perplexity.html). However this can potentially be included in the library with the computation being done in a batched manner.

## Motivation

This would make it easier and faster for people to evaluate their language models in terms of perplexity.

If not a solution integrated in the library, the example given in the docs can be updated to do computation in a batched manner for speed.