server: advertise loaded vision input capability

[apetersson]

Motivation

Clients should be able to discover whether the currently loaded model accepts images instead of treating every DS4 model as text-only.

Summary

• add an input field to /v1/models entries
• advertise ["text", "image"] only when the loaded engine has vision
• advertise ["text"] for text-only engines
• cover both cases in the server metadata regression test

This lets OpenAI-compatible clients discover image support from the model endpoint instead of assuming the server is text-only.

Testing

• make -j ds4-server ds4_test
• ./ds4_test --server
• git diff --check upstream/main...HEAD