Add ImageProcessorFast to Efficientnet processor

[Yann-CV]

What does this PR do?

Following #36978:
This pull request introduces a new fast image processor for EfficientNet models and integrates it into the existing codebase. The changes include updates to documentation, initialization files, and test cases to support the new EfficientNetImageProcessorFast.

Integration of EfficientNetImageProcessorFast:

• docs/source/en/model_doc/efficientnet.md: Added documentation for EfficientNetImageProcessorFast.
• src/transformers/__init__.py: Included EfficientNetImageProcessorFast in the import structure and import statements. [1] [2]
• src/transformers/models/auto/image_processing_auto.py: Updated the image processor mapping to include EfficientNetImageProcessorFast. [1] [2]
• src/transformers/models/efficientnet/__init__.py: Added import for EfficientNetImageProcessorFast.

Implementation of EfficientNetImageProcessorFast:

• src/transformers/models/efficientnet/image_processing_efficientnet_fast.py: Added the implementation of the EfficientNetImageProcessorFast class, including methods for preprocessing, rescaling, and normalizing images.

Testing and Dummy Objects:

• src/transformers/utils/dummy_torchvision_objects.py: Added a dummy class for EfficientNetImageProcessorFast to handle cases where torchvision is not available.
• tests/models/efficientnet/test_image_processing_efficientnet.py: Updated test cases to include EfficientNetImageProcessorFast and ensure it is tested alongside the standard EfficientNetImageProcessor. [1] [2] [3]

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[Yann-CV]

I am spotting an inconsistency between slow (in efficientnet) and fast resize (base class):

    def test_resize(self):
        torch.manual_seed(0)
        data =  torch.randint(0, 256, (1, 3, 480, 640), dtype=torch.uint8)
        for image_processor_class in self.image_processor_list:
            image_processor = image_processor_class(**self.image_processor_dict)
            if image_processor_class == EfficientNetImageProcessorFast:
                image = data
                print(f"image: {image.shape} - {image.mean(dtype=torch.float32)}")
                resized_image = image_processor.resize(
                    image,
                    size=SizeDict(height=18, width=18),
                    interpolation=F.InterpolationMode.NEAREST,
                    antialias=False,
                )
                print(f"resized_image: {resized_image.shape} - {resized_image.mean(dtype=torch.float32)}")
            else:
                image = data.squeeze().permute(1, 2, 0).numpy()
                print(f"image: {image.shape} - {image.mean()}")
                resized_image = image_processor.resize(
                    image,
                    size={"height": 18, "width": 18},
                    resample=PILImageResampling.NEAREST
                )
                print(f"resized_image: {resized_image.shape} - {resized_image.mean()}")

is returning

image: (480, 640, 3) - 127.5092306857639
resized_image: (18, 18, 3) - 126.91769547325103
image: torch.Size([1, 3, 480, 640]) - 127.50922393798828
resized_image: torch.Size([1, 3, 18, 18]) - 121.97119140625

You can obtain quite a gap in mean value for the same interpolation mode. antialias does not impact the result

[Yann-CV]

All other method are outputing the same result (normalize and rescale) between slow and fast

[Yann-CV]

using bilinear resizing is making the test pass (too much difference between PIL and torchvision with nearest)

[Yann-CV]

@ydshieh @yonigozlan the test are failing on test non related to this pull request.
In addition of updating the resize interpolation in tests, I took the freedom to fix the seed of numpy in the tests.