Add support for VLMs in vLLM

[tengomucho]

What does this PR do?

Enable support for VLMs in vLLM, and add a test to show functionality.

[Copilot]

Pull request overview

Adds initial vision-language model (VLM) support to the Optimum Neuron vLLM integration by introducing a multimodal runner path and extending the vLLM service test client/tests to exercise image-conditioned generation.

Changes:

• Add a VLM-specific vLLM runner that extracts pixel_values from vLLM multimodal features and passes them into Neuron models during prefill.
• Introduce an OptimumNeuronModelForImageTextToText wrapper for vLLM model loading and forward/prefill APIs.
• Extend vLLM service fixtures/client and add a new service test for generation with images.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
tests/vllm/service/test_vllm_service_generate.py	Refactors service fixture creation and adds a VLM image generation service test.
tests/fixtures/llm/vllm_service.py	Extends the OpenAI-compatible test client with image-capable chat sampling helpers.
optimum/neuron/vllm/runner.py	Adds a VLM runner variant and refactors token generation helpers for reuse.
optimum/neuron/vllm/model_loader.py	Adds a VLM model wrapper and introduces a task mapping entry for image-text-to-text.
optimum/neuron/models/inference/backend/modules/decoder/vlm_decoder.py	Exposes a reusable prepare_vlm_prefill() for chunked prefill image feature preparation.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.