guide llms.txt improvements

[thibaudcolas]

Follow-up improvements to #412 / #413, based on further testing of the files and feedback (thread: llms.txt feedback #13648).

• Extract as a package
• Set up CDN-level caching of text/plain and text/markdown responses
• Remove "releases" pages from the files (they shouldn’t be relevant to answer questions about the CMS)
• Track files’ token lengths
• Set up evals suite for guide contents with common questions #415
• Unit tests for lack of HTML escaping
• Trial LLM-focused information
  • Information only present in llms.txt
  • Information only present in pages’ markdown representations
  • Information only present in llms-full.txt
  • Page only linked from llms.txt
• Make the files discoverable directly from the site: Implement Markdown / LLM content copy features #416

To be confirmed

Potential improvements where I’m not sure how high the ROI is:

• TBC: remove "About" and "Contributing" pages to increase signal to noise
• TBC: set a target of token count so the file fits in small models’ context windows (for example to fit in a 32k context)
• TBC: add inline heading links to Markdown to encourage direct linking to relevant sections
• TBC: change the "full" format so there is less ambiguity about documents’ start and end
• TBC: docs-focused MCP server