docs: comprehensive README rewrite based on actual codebase

[superdd-coder]

Summary

Complete rewrite of README.md and README_CN.md based on thorough analysis of the actual codebase.

Changes

• New "How It Works" section — explains the Spark/Sink/Educe philosophy with concrete technical depth: meeting pipeline (WebSocket transcription → two-pass LLM → Blueprint auto-sectioning → per-section summaries), Notes (Distill + Propagate + bidirectional reference graph), document ingestion (chunking, contextual retrieval, consolidation), and retrieval (Direct vs Agentic RAG, session-based SSE chat, recall evaluation)
• Feature tables reorganized — Meetings, Notes, Ingestion & Organization, Retrieval & Reasoning, and Extensibility each get dedicated tables with user-facing descriptions
• Blueprint auto-sectioning highlighted — describes how the LLM uses existing Collection catalog as a classification taxonomy, matches topics to existing Collections and suggests new ones at the same dimension
• MinerU configuration guide — added as a Quick Start TIP with 3-step setup
• DashScope transcription recommendation — noted as the recommended path with OneShot setup
• Educe MCP mention — brief introduction connecting MCP to the Educe philosophy of letting knowledge flow into external tools
• Removed overly technical details from user-facing sections (timing constants, internal file names, API paths, SSE event names, library names, class names, programming patterns)
• Both languages kept semantically identical