Implement Friendli AI adapter

[ZHOUEU-star]

Summary

• replace the Friendli stub with a Serverless OpenAI-compatible chat completions integration
• use FRIENDLI_TOKEN and the documented serverless chat completions endpoint
• support model overrides, system prompts, max tokens, temperature, extra request fields, dry-run behavior, and usage token mapping

Verification

• pnpm --filter @profullstack/sh1pt-ai-friendli typecheck
• tsx dry-run smoke check for missing-key and dry-run paths

Note: local vitest startup is blocked by a macOS Rollup native package signature loading error before tests execute.