feat: add Stability AI SD3.5 provider for permissive image generation

[realityinspector]

Summary

• Adds Stability AI as a new provider for image generation using the SD3.5 Large model via REST API
• Enables the permissive/free-distillable pipeline to generate images (previously text-only)
• SD3.5 Large allows downstream distillation, fitting the permissive model policy
• Includes fallback to OpenRouter (FLUX) if Stability API fails

Changes

• New file: app/core/providers/stability.py — Stability AI REST API client (multipart/form-data to v2beta/stable-image/generate/sd3)
• app/config.py — Add STABILITY to ProviderType, STABILITY_API_KEY to Settings, STABILITY_IMAGE to VerifiedModels, update FREE_DISTILLABLE preset with image_model: "stability-ai/sd3.5-large"
• app/core/model_policy.py — Add stability-ai/ to PERMISSIVE_PREFIXES, handle in derive_model_provider
• app/core/model_capabilities.py — Add STABILITY to ImageModelType, register stability-ai/sd3.5-large in image model registry, add is_stability_model() helper
• app/core/providers/__init__.py — Export StabilityProvider
• app/core/llm_router.py — Initialize Stability provider, route permissive mode to Stability for images, fallback to OpenRouter FLUX on failure

Configuration

Requires STABILITY_API_KEY env var (already set on Railway).

Test plan

• Verify STABILITY_API_KEY is set in Railway environment
• Test FREE_DISTILLABLE preset generates images via Stability API
• Test permissive mode routes to Stability for image generation
• Test fallback to OpenRouter FLUX when Stability is unavailable
• Verify non-permissive presets (HD, BALANCED) still use Google as before