fix: normalize image URL content type

[z2z23n0]

What changed

• Normalize message content items from type: "image_url" to the model-supported type: "image" before offline inference.
• Preserve the existing image_url value and remote/local path handling.
• Add regression coverage for the normalized model input.

Why

The inference wrapper treated type: "image_url" content as image input, but the model only emits an image token for type: "image" or items with an image field. As a result, an image_url item could reach the model without a vision token and silently run as text-only inference.

Validation

• python3 -m unittest inference.tests.test_run_inference
• uvx ruff check inference/run_inference.py inference/tests/test_run_inference.py
• git diff --check