[v1] Add PrefixLM support to FlexAttention backend

[Isotr0py]

Purpose

• Currently, there is no attention backend supports image-bidirectional attention in vLLM, so Gemma3 and paligemma can't generate correct outputs. And models like moondream are blocked due to missing attention backend support.
• This PR adds image-bidirectional support to FlexAttention backend to fill the void.

Test Plan

pytest -s -v tests/models/multimodal/generation/test_common.py -k gemma3

Test Result

vllm (both native and transformers backend) results should have converged results with HF now.

PR results

tests/models/multimodal/generation/test_common.py::test_single_image_models[gemma3-test_case0]
  /home/mozf/develop-projects/vllm/tests/models/multimodal/generation/vlm_utils/core.py:157: UserWarning: Test0:
  hf:   "Here's what's in the center of the image:\n\nIt's a traditional Chinese gate or archway. It's red and gold, with Chinese characters written on it. It's a prominent feature of the street scene.<end_of_turn>"
  vllm: "Here's what's in the center of the image:\n\nIt's a traditional Chinese gate or archway. It's red and gold, with Chinese characters written on it. It's a prominent feature of the street scene."
    comparator(

tests/models/multimodal/generation/test_common.py::test_single_image_models[gemma3-test_case0]
  /home/mozf/develop-projects/vllm/tests/models/multimodal/generation/vlm_utils/core.py:157: UserWarning: Test1:
  hf:   'The center of the image features a vibrant Chinese-themed archway or gate. It\'s decorated with red and gold colors, traditional Chinese characters (likely meaning "Chinese Town"), and red lanterns.  There are also two white stone lion statues flanking the entrance.<end_of_turn>'
  vllm: 'The center of the image features a vibrant Chinese-themed archway or gate. It\'s decorated with red and gold colors, traditional Chinese characters (likely meaning "Chinese Town"), and red lanterns.  There are also two white stone lion statues flanking the entrance.'
    comparator(

Main branch

tests/models/multimodal/generation/test_common.py::test_single_image_models[gemma3-test_case0]
  /home/mozf/develop-projects/vllm/tests/models/multimodal/generation/vlm_utils/core.py:157: UserWarning: Test0:
  Matched tokens:       [8291, 236789, 236751, 1144, 236789, 236751, 528, 506, 3988, 529, 506, 2471, 236787, 108]
  hf:   "Here's what's in the center of the image:\n\nIt's a traditional Chinese gate or archway. It's red and gold, with Chinese characters written on it. It's a prominent feature of the street scene.<end_of_turn>"      {1509: -0.03537141531705856, 236776: -3.7853713035583496, 818: -4.53537130355835, 3810: -6.78537130355835, 236829: -8.535371780395508}
  vllm: 'Here\'s what\'s in the center of the image:\n\nA vibrant Chinese-themed archway with red and gold decorations, featuring the Chinese characters "中华" (Zhōnghuá - meaning "China"). It\'s part of a Chinatown area.'       {236776: Logprob(logprob=-0.6554893255233765, rank=1, decoded_token='A'), 818: Logprob(logprob=-1.1554893255233765, rank=2, decoded_token='The'), 1509: Logprob(logprob=-2.155489444732666, rank=3, decoded_token='It'), 236829: Logprob(logprob=-3.155489444732666, rank=4, decoded_token='*'), 3810: Logprob(logprob=-4.905489444732666, rank=5, decoded_token='There')}
    comparator(

tests/models/multimodal/generation/test_common.py::test_single_image_models[gemma3-test_case0]
  /home/mozf/develop-projects/vllm/tests/models/multimodal/generation/vlm_utils/core.py:157: UserWarning: Test1:
  Matched tokens:       []
  hf:   'The center of the image features a vibrant Chinese-themed archway or gate. It\'s decorated with red and gold colors, traditional Chinese characters (likely meaning "Chinese Town"), and red lanterns.  There are also two white stone lion statues flanking the entrance.<end_of_turn>'    {818: -0.5989671945571899, 8291: -0.8489671945571899, 117494: -4.5989670753479, 6481: -5.3489670753479, 19058: -5.5989670753479}
  vllm: 'Here\'s what\'s in the center of the image:\n\n*   **A large, ornate Chinese gate or archway.** It\'s painted red and features traditional Chinese characters ("中华" - meaning "China") and decorative elements.\n*   **Two white stone lion statues** flanking the gate.\n*   **A black SUV** is parked in front of the gate.\n\nLet me know if you want me to describe any other specific elements in the image!'     {8291: Logprob(logprob=-0.3420378267765045, rank=1, decoded_token='Here'), 818: Logprob(logprob=-1.3420377969741821, rank=2, decoded_token='The'), 117494: Logprob(logprob=-4.342037677764893, rank=3, decoded_token='Certainly'), 6481: Logprob(logprob=-4.842037677764893, rank=4, decoded_token='Let'), 19058: Logprob(logprob=-5.592037677764893, rank=5, decoded_token='Okay')}
    comparator(

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[mergify]

Documentation preview: https://vllm--27938.org.readthedocs.build/en/27938/

[DarkLight1337]

Let's get this merged then, can you fix the merge conflicts?

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @Isotr0py.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork