Feedback welcome, notably for the design, code quality, etc.

This PR aims at introducing an unified way to benchmark transformers vs. optimized models, backend-independent (in the sense of, any backend can be plugged for inference and evaluation), code-free (in the sense of, the user does not need to code to start runs and evaluate them).

The two main contributions is to introduce helper classes, methods for data preprocessing, inference, evaluation. In several files:
* optimum/runs_base.py: general methods, this should be backend-agnostic.
* optimum/utils/preprocessing/: handle loading and preprocessing datasets, running inference with pipelines, running evaluation. This should be backend-agnostic.
* optimum/onnxruntime/runs/: OnnxRuntime specific methods

For now, dataset preprocessing and evaluation are task-specific, the supported tasks are:

• text-classification
• token-classification
• question-answering

As for evaluation of transformers models, I believe there is some duplicate work with what exists in the AutoTrain backend and what is being done in https://github.com/huggingface/evaluate. However, my understanding being that it is not a priority to support Optimum-based inference within AutoTrain, it makes sense to me to have a common implementation to evaluate transformers/optimized models for them to be comparable. I hope we can make it such that we minimize duplicate efforts.

I used pipelines for inference for the general metrics, and ORTModel.forward() to measure latency/throughput.

Tasks before (or after) merge

• Documentation (Added in an "API Reference" section)
• Test on several datasets
• See if it would make sense to use Trainer.evaluate() instead of an explicit loop for evaluation --> I think it doesn't, there is a lot of abstraction in pipelines already, we should make use of it.
• Make use of train-eval-index metadata from datasets to auto-infer data, label columns (see e.g. Autoeval config datasets#4234) (left to next PR)
• Support multi-column data (2 would be sufficient I guess, see Bert that receives text triplet as an input transformers#8573)
• Document node exclusion (left to next PR)
• Support node exclusion for dynamic quantization for OnnxRuntime (implemented in Allow onnxruntime quantization preprocessor for dynamic quantization #196)
• Avoid tracking PyTorch to Numpy conversion for time measurements (left to next PR), in

  optimum/optimum/onnxruntime/modeling_ort.py

  Lines 323 to 327 in cf91bd7

  	# converts pytorch inputs into numpy inputs for onnx
  	onnx_inputs = {
  	"input_ids": input_ids.cpu().detach().numpy(),
  	"attention_mask": attention_mask.cpu().detach().numpy(),
  	}

• Still some work to distinguish backend-agnostic code vs. backend-specific code
• Clean all remainings # TODO
• Unit tests / github workflows (left to next PR)

This, with some additional work, should close #128 .