Match document language in LLM image descriptions

[Br1an67]

Summary

Fix LLM-generated image descriptions ignoring document language — e.g., producing English descriptions for French PDFs.

Two changes:

1. Added explicit language-matching instruction (rule Poor parsing #4) to the image description prompt, telling the LLM to output descriptions in the same language as the input text
2. When the image itself has little/no text (< 20 chars), surrounding page text (up to 200 chars) is included as language context, so the LLM can detect the document language even for pure diagrams

Closes #954

Changes

• Updated image_description_prompt in LLMImageDescriptionProcessor with language-matching instruction
• Modified block_prompts() to include surrounding page text for language context when the image's own text is sparse