[Patch] Add loss for ORT inference

[JingyaHuang]

What does this PR do?

• Wrap OnnxConfig by wrap_onnx_config_for_loss to obtain the loss while using ORTTrainer under the mode inference_with_ort=True.
• Enable deepspeed for ONNX Runtime training. (Tested with ZeRO stage 2, full availability under progress)
• Clean up unused dependencies in ORTTrainer.
• Update CI of onnxruntime training.
• Update associated tests.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[lewtun]

This line is why the doc build is currently failing: https://github.com/huggingface/optimum/pull/152/files#diff-3e928ce0b52f617b86cd0df9399c6bbb5804d6269c6a04b486613eb929449256R25

Until now we never actually imported OnnxConfigWithLoss etc and doing so triggers an import error because transformers is pinned to <4.17 in the doc build (to have both intel and onnxruntime packages in the same env):

from optimum.onnxruntime.trainer import ORTTrainer

ImportError: cannot import name 'TensorType' from 'transformers.utils' (/home/lewis/miniconda3/envs/optimum/lib/python3.8/site-packages/transformers/utils/__init__.py)

The error arises because TensorType was moved from transformers.file_utils to transformers.utils in >=4.17

The solution is to refactor the doc build so that we build intel and onnxruntime in separate envs. Doing so will also allow us to build the Graphcore & other hardware partner docs as well. I don't have bandwidth for this right now, but happy to review a PR if someone else has time to tackle this!

[lewtun]

FYI @JingyaHuang if you want to test that the docs build locally you can run:

pip install -e '.[dev,intel.onnxruntime]'
pip install git+https://github.com/huggingface/doc-builder.git
doc-builder build optimum docs/source --build_dir test-docs --version v1.0.0 --clean

You'll need a Linux machine for this since one cannot install intel on macOS

[echarlaix]

This PR looks great !