What does this PR do?

TL;DR Preprocessing and postprocessing ops can be offloaded to Hub kernels the same way model layers currently are. This PR allows this and goes with a first simple kernel that is common to many image processors.

use_kernels=True swaps a processor op for a registered Hub kernel and falls back to the default implementation when none applies. So custom or hard-to-vectorize ops (beyond resize or normalize: patchify, NMS, mask decoding, ...) can move onto the GPU without rewriting the processor or the call site.

However the adapter is uglier than for modeling components, mainly because there's no same-signature forward to swap and no kernelize pass over processors. Open to ideas here!

I pushed a fast image processors' resize + normalize that is automatic for processors using the base fast-processing path others fall back and can opt in later. The kernel batches the resize over a ragged set of images and folds rescale and normalize into the same pass.

from transformers import AutoImageProcessor

processor = AutoImageProcessor.from_pretrained("google/siglip2-base-patch16-224", use_kernels=True)
pixel_values = processor(images, return_tensors="pt")["pixel_values"]

For vision-language models the image processor is wrapped in a processor, so the flag passes through [AutoProcessor] to the inner image processor.

from transformers import AutoProcessor

processor = AutoProcessor.from_pretrained("google/gemma-3-4b-it", use_kernels=True)
inputs = processor(text=prompt, images=images, return_tensors="pt")

Some benchmarks for processors of vision models showing the speedup:

processor() resize+normalize, use_kernels off -> on:
siglip2-base-patch16-224     3.92 -> 1.30 ms   3.0x   pixΔ 3.9e-3
paligemma2-3b-pt-224         4.26 -> 1.36 ms   3.1x   pixΔ 3.9e-3
clip-vit-base-patch32        7.30 -> 1.41 ms   5.2x   pixΔ 1.8e-2
dinov2-small                 7.26 -> 1.42 ms   5.1x   pixΔ 2.5e-1


standalone resize stage (siglip-so400m, bicubic+AA):
torchvision eager loop   2.95 ms
torch.compile (dynamic)  5.65 ms   (slower; ~2.6s warmup)
fused triton (2D)       11.53 ms
separable triton        1.37 ms    -> 2.9x the real processor (3.96 ms)

Need to move the kernel from Molbap/<kernel_name> and open this to the community. Feedback welcome :)