fix: add OpenRouter image generation support & fix retry task directory

[voidborne-d]

Problem

1. OpenRouter API incompatibility (image_api.py)

When using OpenRouter as the image provider with endpoint_type: chat, the Chat API generator:

• Does not include modalities: ["image", "text"] and image_config in the request payload, so OpenRouter returns text-only responses
• Does not parse OpenRouter's message.images[] array format, only looking for images in message.content

2. Retry fails with "任务目录未设置" (image.py)

retry_failed_images() never sets self.current_task_dir before spawning worker threads. When workers call _generate_single_image() → _save_image(), the task directory is None, causing all retries to fail even though image generation succeeds.

Changes

backend/generators/image_api.py

• Add modalities: ["image", "text"] and image_config (with aspect_ratio and image_size) to Chat API request payload
• Parse OpenRouter's message.images[] array response format (base64 data URLs)
• Handle multimodal content parts (content as list of image_url objects)
• Preserves backward compatibility with existing response formats (markdown URLs, plain base64, etc.)

backend/services/image.py

• Set self.current_task_dir at the start of retry_failed_images(), consistent with generate_images() and retry_single_image()

OpenRouter Config Example

active_provider: openrouter_gemini

providers:
  openrouter_gemini:
    type: image_api
    api_key: sk-or-v1-xxxxx
    base_url: https://openrouter.ai/api/v1
    model: google/gemini-3.1-flash-image-preview
    endpoint_type: chat
    high_concurrency: false
    image_size: 4K

Testing

• Tested with google/gemini-3.1-flash-image-preview via OpenRouter
• Successfully generated 7-page Xiaohongshu post (cover + 5 content + summary)
• Retry ("一键补全失败图片") works correctly after fix
• Existing images/generations endpoint path still works (no regression)