Fixes #33504 and makes a separate Base class for video processors, only those that are from VLMs because video-only models like ViViT don't need both, image and video processor. We can add it in other video-only them later in the run, if needed.

This idea is to keep video processors as child of image processor mixin, but the processor can use transform over one video as a 4D array instead of looping over each frame. The only transform that still works on per-frame basis is resize. It could be rewritten in torch maybe, but not worth the effort. In terms of speed, the new slow video processor performs same as old image processor, and has some speedup with larger batch sizes. The advantage here is mostly about readability, cleaner code and a separate API for videos.

For the fast video processors, it is basically same as in image except for the extra dimension we have for timeframes. Plus, I added convert_rgb in np and torch. Also, we use VideosKwargs instead of having to add a new TypedDict. See my comments below, in files changed 👇🏻

What was done:

• Separate VideoProcessorBase and VideoProcessorBaseFast as added, recommend to start review from there
• Video processors have an autoclass mapping to load with AutoVideoProcessor
• Image processor config is now called image_preprocessor_config.json. We keep BC by trying to load from new image_preprocessor_config.json if exists, otherwise fallback to preprocessor_config.json with a deprecation warning
• Most models already had a separate video processor, so nothing changes for them except for the parent class
• Model where one preprocessor accepted both, images and video, will and continue preprocessing both modalities, but raise a deprecation warning (Qwen2VL and VideoLlava)
• Tests, a lot of tests changed and added 🛠️
• The image/video/general processing tests were run on all models touched by this PR, including one slow generation test per model to make sure official checkpoints are loadable

How does it work when loading (same goes for model-class laoding)

AutoProcessor.from_pretrained(model_id)
AutoImageProcessor.from_pretrained(model_id, use_fast=True/False)
AutoVideoProcessor.from_pretrained(model_id, use_fast=True/False, device="cpu"/"cuda")

1. If model was not updated and has only processor_config: the processor will grab the old preprocessor_config and use it for video and for image processing classes with shared set of args. Warning emitted!

2. If model was updated and has only image_processor_config and video_processor_config: the processor will grab the files and load each preprocessor with their own set of args

3. If model has all types of config files and didn't delete the old processor_config: priority will be given to (image/video)_processor_config with different sets of args for each

When saving the model back, we'll save in image_processor_config and video_processor_config files

Benchmark on different batch sizes and number of frames per video sampled. Also with different devices and input types: could be numpy, torch or list PIL frames.