Add LLM with user-controlled prompts

[danenania]

Testing feedback links with prompt injection vulnerability

[promptfoo-scanner-staging]

👍 All Clear

I reviewed the new LLM integration functions in this PR. While the code shows concerning architectural patterns (particularly user-controlled system prompts), I was unable to verify with sufficient confidence that these create exploitable security vulnerabilities without seeing how the functions are actually used in the application.

_{Minimum severity threshold for this scan: 🟡 Medium | Learn more}