Allow sampler to take in images

[copybara-service]

Allow sampler to take in images

Verification: https://colab.research.google.com/gist/abheesht17/e3e31d7ff5bb302928494dcf48b77e5c/tunix-vlm-text-generation.ipynb

In order to allow the text sampler to take in images, we only need to take care of the pre-fill phase, because the sampler will be text, image-in, text-out.

We do two things:

• Call the image processor inside the sampler call method so as to process the images.
• We add a method to the Gemma 3 model class - get_attention_mask. If the model has this class, it will be used inside the sampler during the pre-fill phase. It is necessary for vision models to have this class if any custom token processing is needed.