feat: add local AI image generation via ComfyUI (replacing Pexels)

[elulue]

Summary

This PR adds a new video_source="ai" option that generates video materials locally using ComfyUI, instead of searching Pexels/Pixabay stock footage.

• New file: app/services/comfyui.py — ComfyUI API client with FLUX.1-schnell text-to-image workflow + Ken Burns effect (zoom/pan) for automatic image-to-video conversion
• Modified: app/services/task.py — added elif video_source == "ai" branch in get_video_materials()
• Modified: app/models/schema.py — documented "ai" as valid video_source value
• Modified: config.example.toml — added ComfyUI configuration section

How it works

search_terms → enhance_prompt() → ComfyUI FLUX.1-schnell generate image
→ FFmpeg Ken Burns effect (random zoom-in/zoom-out/pan) → 5s video clip
→ repeat until audio duration is covered → return video paths (same interface as Pexels)

Benefits

• Fully local generation, zero API costs
• Zero copyright concerns (all AI-generated)
• Content precisely matches the script (no keyword-to-stock-footage mismatch)
• Works with any ComfyUI-supported image model (FLUX, SD3, etc.)
• Configurable prompt prefix/suffix for style control
• Non-breaking: existing Pexels/Pixabay workflows remain unchanged

Configuration

Add to config.toml:

comfyui_api_url = "http://127.0.0.1:8188"
comfyui_flux_model = "flux1-schnell.safetensors"
comfyui_flux_vae = "ae.safetensors"
comfyui_t5xxl = "t5xxl_fp16.safetensors"
comfyui_clip_l = "clip_l.safetensors"
comfyui_timeout = 120
comfyui_prompt_prefix = "Professional cinematic photo, high quality, detailed, 4K, "

Then set video_source = "ai" in WebUI or API request.

Requirements

• A running ComfyUI instance with FLUX.1-schnell (or compatible) model loaded
• FFmpeg installed (for Ken Burns effect conversion)

Closes #586