NSP Support for Zero-shot Text Classification Pipeline

### Feature request Zero-shot classification can be solved with NextSentencePrediction task of BERT, and it has shown competitive results to NLI-based zero-shot classification in some cases. There could be a parameter where we choose the type of submethod that we are going to use for the pipeline like `pipeline(task="zero-shot-classification", type_="nsp")` or we could just simply add a task named "nsp-zeroshot-classification". This is also possible for MLM, which is a more widely used pretraining task across LMs. ### Motivation Like I said, NSP has proven to be useful especially for languages that do not have access to NLI dataset since only pre-training is enough. Although multilingual NLI models can also be used, they have been proven to be worse compared to smaller monolingual models in this task, as one would expect. Even if this is a small detail which would be unnecessary to put into the codebase, I wanted to share this implementation so that anyone who's interested can take a look and try different methods. Here are some references, one of which is my study, that use NSP for zero-shot classification. Sun, Y., Zheng, Y., Hao, C., & Qiu, H. (2021). NSP-BERT: A Prompt-based Zero-Shot Learner Through an Original Pre-training Task--Next Sentence Prediction. arXiv preprint arXiv:2109.03564. Çelik, E., & Dalyan, T. (2023). Unified benchmark for zero-shot Turkish text classification. Information Processing & Management, 60(3), 103298. ### Your contribution I can open a PR, here's the implementation I did based on Sun et al. 2021. It is heaily based on the current NLI zeroshot pipeline class, but also adds a `reverse` argument which changes the order of the sentences for NSP. ```python import numpy as np from typing import List, Union from transformers.utils import logging from transformers.pipelines.base import ChunkPipeline, ArgumentHandler from transformers.tokenization_utils import TruncationStrategy from transformers.pipelines import ZeroShotClassificationArgume…