fix(offload): preserve namespaced model ids

[YOMXXX]

Summary

• Fixes [Feature] 希望offload modelId兼容 类似 siliconflow/deepseek-ai/DeepSeek-V4-Flash 这样的id #24 by parsing provider/model refs at the first / only, so model ids like deepseek-ai/DeepSeek-V4-Flash are preserved.
• Reuses the same parser for offload local LLM setup and context-window lookup.
• Moves the existing clean-context model-ref helpers into a lightweight shared utility and keeps the old exports for compatibility.
• Adds a regression test that exercises registerOffload() and verifies the initialized local LLM model id is not truncated.

This is a fresh, mergeable replacement for the older conflicting PR #33.

Verification

• npx vitest run src/offload/model-ref.test.ts
• npm test (4 files, 48 tests)
• npm run build
• git diff --check

[Maxwell-Code07]

Awesome work! 13 fixes in one batch covering Gateway hardening, embedding compatibility, Hermes memory injection, offload stability, and pipeline optimization — a very productive cleanup session 👏

We'll review and merge them one by one 🚀