fix: use matching Undici fetch for private AI providers

[singhangadin]

Overview

Fix allowlisted private AI providers failing with fetch failed when Termix runs with a different bundled Undici version than its installed undici package.

• Fixed: Private AI provider requests now use the installed Undici fetch implementation with its matching dispatcher.
• Added: Regression coverage for the private-provider request path.

Changes Made

• Route allowlisted private provider requests through the existing fetchWithProxy helper.
• Preserve proxy handling and the existing private-endpoint allowlist checks.
• Avoid passing an Undici 8 dispatcher to Node's bundled Undici 7 fetch, which fails with UND_ERR_INVALID_ARG: invalid onRequestStart method.

Related Issues

• Closes AI provider requests to private/allowlisted hosts fail with generic "fetch failed" (undici dispatcher ABI mismatch — same class as #1185, not covered by #1291's fix) #1298

Screenshots / Demos

• I tested this fix successfully on my self-hosted Termix 2.7.1 Docker deployment with Ollama on an allowlisted private endpoint. Model discovery and streamed chat both work after the fix.

Testing

• 26 focused backend tests passed.
• TypeScript project check passed.
• ESLint passed for the changed files.

Checklist

• Code follows project style guidelines
• Supports mobile and desktop UI/app (not applicable; backend-only change)
• I have read Contributing.md
• This is not a translation request. See docs

[ZacharyZcR]

改动范围聚焦，已在 dev-2.8.0 上复核；完整 CI 与数据库方言测试均通过。