docs(extension): plan browser-local Gemma 4 E4B

[BjoernSchotte]

Summary

• add the architecture and implementation plan for optional browser-local Gemma 4 E4B inference with Transformers.js and ONNX Runtime WebGPU
• preserve Anthropic as the existing first-class provider and scope Gemma to Quick, Auto, and Think deeper while excluding Deep Research
• structure delivery around an early production-path GO/NO-GO slice through the existing extension Chat system, followed by provider, model lifecycle, runtime, UX, quality, privacy, and documentation hardening
• define a provider-ready UI contract without implementing custom or OpenAI-compatible endpoints in this scope

Why

The extension needs a proof-gated path to fully local LLM inference without introducing a separate Chat implementation or weakening the existing agent, tool, answer, citation, persistence, and privacy contracts.

Impact

This PR contains planning documentation only. It does not add a runtime dependency or change extension behavior.

Validation

• bun run docs:check
• bun run typecheck
• staged diff and whitespace validation