feat: Improve image generator tuning for manga panels

[datnguyen1]

Summary

Improve the quality and consistency of AI-generated comic panel images.

Current State

• Uses FLUX.1-schnell with a style prompt (shounen, shoujo, etc.) + panel description
• Single image model; no negative prompts or parameter tuning

Proposed Improvements

• Add negative prompts (e.g. "no text, no speech bubbles, no watermarks")
• Expose/tune generation parameters (steps, guidance scale, aspect ratio) if supported by the provider
• Refine style prompts per genre for better manga aesthetics
• Consider aspect ratio (e.g. vertical/portrait for comic panels)
• Add retry logic or fallback if image generation fails
• Optional: Support alternative image models for comparison

Notes

• Logic lives in backend/src/services/aiService.js (STYLE_PROMPTS, hfImage, generatePanelImage)