Add image height and width to ONNX dynamic axes

[lewtun]

What does this PR do?

This PR enables dynamic axes for image height / width of ONNX vision models. This allows users to change the height and width of their inputs at runtime with values different from those used to trace the model during the export (usually 224 x 224 pixels)

Here's an example with ResNet and optimum:

import requests
from PIL import Image
from optimum.onnxruntime import ORTModelForImageClassification
from transformers import AutoFeatureExtractor

url = "http://images.cocodataset.org/val2017/000000039769.jpg"
# Raw image size 480 x 640 pixels
image = Image.open(requests.get(url, stream=True).raw)
# Resize image to 40 x 40 pixels
preprocessor = AutoFeatureExtractor.from_pretrained("microsoft/resnet-50",  do_resize=True, size=40)
model = ORTModelForImageClassification.from_pretrained("onnx")
inputs = preprocessor(images=image, return_tensors="pt")
outputs = model(**inputs)
logits = outputs.logits
logits.shape

I've also checked the slow tests pass:

RUN_SLOW=1 pytest tests/onnx/test_onnx_v2.py -k "beit or clip or convnext or data2vec-vision or deit or detr or layoutlmv3 or levit or mobilevit or resnet or vit" -s

[lewtun]

I noticed that the CLIP export was also missing num_channels as a dynamic axis, so included it here as well

[lewtun]

Following #17976 I've renamed sequence to num_channels

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[regisss]

LGTM thanks @lewtun!!

[sgugger]

LGTM, thanks for working on this!