fix HunyuanOCR crash in vLLM

[souvikchand]

This PR fixes a runtime error in the vLLM multimodal pipeline when running HunyuanOCR.
The issue #35 was caused by sending images using the wrong message schema, which led vLLM to misinterpret the image input and generate an invalid tensor shape.

ValueError: image_grid_thw has rank 3 but expected 2.
Expected shape: ('ni', 3), but got torch.Size([2, 1, 3])

what i changed

1. updated request format

{ "type": "image_url", "image_url": { "url": "data:image/jpeg;base64,..." } }

to

{
    "type": "image_url",
    "image_url": f"data:{mime};base64,{encode_image(image_path)}"
 },

2. Added automatic MIME-type detection to ensure images are sent with the correct format (png/jpeg/webp/etc)
3. Ensured image_url is a string, not a nested object, which aligns with vLLM’s expected schema for HuggingFace vision models.

[diegocarturan-debug]

[souvikchand]

@diegocarturan-debug
sorry but can you explain your comment above. it's actually redirecting to this same page