[Feature] Route slide/figure image generation through Kady's configured providers instead of hardcoded OpenRouter

[arsaboo]

The scientific-slides skill generates presentation visuals (and full-slide images for the PDF workflow) via generate_slide_image.py / generate_slide_image_ai.py. Those scripts currently hardcode https://openrouter.ai/api/v1 and call google/gemini-3.1-flash-image ("Nano Banana Pro"), with a gemini-3.7-flash review step, and abort (Error: OPENROUTER_API_KEY not found) when no OpenRouter key is present. There's no fallback to any other provider — not Ollama, NVIDIA, or the configured OAuth providers (OpenAI Codex, xAI, Anthropic).

This hardcodes a dependency on OpenRouter for the default slide workflow, even in a BYOK app that otherwise lets users bring their own model access (OpenAI, xAI, NVIDIA NIM, Ollama). Since these scripts run as sandbox subprocesses, they only see process.env — so they can reach OpenRouter (env key) but not OAuth providers (tokens live in Pi's auth store and are correctly never exported into the sandbox).

Here's what I propose (not yet implemented):

• Add a backend-owned, provider-agnostic image-generation capability (a tool + localhost endpoint, reusing the existing kady-modal/notebook/latex-assist patterns), so credentials stay centralized, budget-gated, and cost-ledgered — and never leak into the sandbox.
• A provider resolver picks an actually-available image model with a sane fallback order (configured image provider → OpenRouter).
• Point the skill scripts at that endpoint instead of the hardcoded OpenRouter URL, making OpenRouter optional.

@TKassis let me know what you think about this.