feat: add image-to-image generation via reference image attachment

[tomasmach]

Summary

• Adds use_reference_image tool parameter so the LLM can signal image-to-image generation when the user attaches an image
• New img2img_model config field (global + per-agent) selects the fal.ai endpoint (fal-ai/flux/dev/image-to-image by default)
• Agent now extracts the first downloaded image attachment as the reference URL and passes it into ImageGenDeps before building the tool registry
• Web API (GET/PUT /api/config/image) and dashboard UI expose the new img2img_model field

Test plan

• All existing tests pass (go test -tags sqlite_fts5 ./...)
• TestImageGenImg2ImgWithReference — verifies img2img endpoint, image_url, strength=0.85, steps=28
• TestImageGenImg2ImgWithoutReference — verifies graceful fallback to text-to-image when no reference URL is available
• TestImageGenNoReferenceFlag — verifies text-to-image is used when flag is false even if reference URL is present
• Manual: attach an image in Discord, ask "generate an image like this but as anime" → bot calls generate_image with use_reference_image: true