feat: LLM readability layer — llms.txt, serve-markdown, MCP discovery, and CDN caching

[Nonnyjoe]

Summary

• Syncs llms.txt to cover all sitemap URLs with correct .md links and AI agent instructions
• Adds skill.md, MCP discovery endpoint (/.well-known/mcp), and a visually-hidden llms.txt link in every page body
• Adds plugins/serve-markdown.js — serves raw .md at build time alongside HTML, handles generated-index pages, and supports Accept: text/markdown content negotiation in the dev server
• Fixes broken internal links across several versioned doc pages
• Adds per-file-type CDN cache headers (customHttp.yml) and Amplify rewrite rules (amplify.yml) so .md files, llms-full.txt, skill.md, and the MCP endpoint are served correctly without being intercepted by the SPA fallback

What is NOT in this PR

The Amplify Gen 2 CDK backend (CloudFront distribution + Lambda function) lives in a separate PR (feat/amplify-lambda) that will be monitored on staging before merging.

Test plan

• https://docs.cartesi.io/llms.txt — all links end with .md and resolve
• https://docs.cartesi.io/cartesi-rollups/overview.md — returns raw markdown directly
• https://docs.cartesi.io/skill.md — resolves
• https://docs.cartesi.io/.well-known/mcp — returns JSON
• curl -H "Accept: text/markdown" https://docs.cartesi.io/cartesi-rollups/overview — returns Content-Type: text/markdown
• HTML pages load normally in browser (no regressions)
• Cache-Control headers correct per file type (assets immutable, html/md short TTL)

Related to #328