[Grounding DINO] Add support for cross-attention in GroundingDinoMultiHeadAttention

[EduardoPach]

What does this PR do?

This PR fixes #30176 by adding support for cross-attention masking and related integration test.

Now one can do batched inference with different text lengths such as:

import torch
from transformers import AutoModelForZeroShotObjectDetection, AutoProcessor

model_id = "IDEA-Research/grounding-dino-tiny"

model = AutoModelForZeroShotObjectDetection.from_pretrained(model_id)
processor = AutoProcessor.from_pretrained(model_id)

images = [...]
text = ["a cat. a remote control.", "a person. a cat. a remote control. a couch."]

inputs = processor(images=image, text=text, padding="longest", return_tensors="pt")

with torch.no_grad():
    outputs = model(**inputs)

[amyeroberts]

Thanks for adding @EduardoPach! Let us know when it's ready for review!

[EduardoPach]

Thanks for adding @EduardoPach! Let us know when it's ready for review!

It's already ready for review, it was easier than I was expecting

[amyeroberts]

Thanks for adding this and a corresponding test! ❤️

Main comments are for robustness when using the model in transformers

[amyeroberts]

Thanks for adding!