Add tabbed content support, URL-level change detection, and more

[djannot]

Three improvements to content processing and crawl reliability, plus E2E test coverage:

• Tabbed content support: Detects WAI-ARIA tabs in web pages and injects tab labels into panel content so context is preserved in the Markdown output instead of concatenating unlabeled code blocks.
• URL-level change detection: Replaces per-chunk hash comparison with URL-level comparison across all source types. Unchanged URLs are skipped entirely; changed URLs are fully re-processed, eliminating orphaned chunks and inconsistent chunk_index/total_chunks when content shifts.
• Puppeteer resilience: Adds protocol and evaluate timeouts, navigates to about:blank between pages, and recreates the page tab after errors to prevent cascading failures from stuck pages.
• E2E integration tests: Three end-to-end tests covering local directory (Markdown), code source (Tree-sitter), and website source (Puppeteer + local HTTP server). Each validates that only modified documents are re-embedded while unchanged ones are skipped, with correct chunk ordering and no orphans.
• Qdrant filter fix: Fixes removeChunksByUrlQdrant using Qdrant's full-text match (match.text) instead of exact keyword match (match.value), which caused cross-URL chunk deletion during content-change re-processing and unnecessary re-embedding of unchanged pages on subsequent runs.
• HTTP 429 retry and browser recovery: Adds rate-limit retry with Retry-After header parsing (max 3 attempts per URL, configurable fallback delay), browser restart escalation when Chrome protocol errors persist after page recreation, periodic browser restart every 50 pages to prevent memory accumulation, and --disable-dev-shm-usage for Docker stability.
• Add ETag-based change detection for website crawls
• Fix shouldProcessUrl rejecting version-like URLs (e.g. /app/2.1.x/)
• Add adaptive HEAD backoff and sitemap lastmod-based change detection
• Testing all 4 change detection layers across multiple syncs