Support image/video in Inference Command

**Is your feature request related to a problem? Please describe:**

Currently matrix inference only support text.

**Describe the solution you would like:**

1. Find huggingface dataset contains images (eg cais/hle) and videos
2. Convert the row into Chat message
3. Deploy maverick-fp8 on h100 and test it on images, which should work.
4. Explore video input and model, eg gemma

**Additional Context:**
We currently integrated vllm v0.8.3, please see if it support video models.